WO2020219991A1 - Medical decision support system - Google Patents

Medical decision support system Download PDF

Info

Publication number
WO2020219991A1
WO2020219991A1 PCT/US2020/029970 US2020029970W WO2020219991A1 WO 2020219991 A1 WO2020219991 A1 WO 2020219991A1 US 2020029970 W US2020029970 W US 2020029970W WO 2020219991 A1 WO2020219991 A1 WO 2020219991A1
Authority
WO
WIPO (PCT)
Prior art keywords
auscultatory
signal
heart
auscultatory sound
sound signal
Prior art date
Application number
PCT/US2020/029970
Other languages
French (fr)
Inventor
Sergey A. Telenkov
Robin F. Castelino
Brian J. Booth
Marina Vernalis
Fatma USTA
Bahareh TAJI
David Gloag
Original Assignee
Ausculsciences, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ausculsciences, Inc. filed Critical Ausculsciences, Inc.
Priority to CA3137910A priority Critical patent/CA3137910A1/en
Priority to EP20794539.5A priority patent/EP3958743A1/en
Publication of WO2020219991A1 publication Critical patent/WO2020219991A1/en
Priority to US17/509,018 priority patent/US20220061797A1/en

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A61B5/349—Detecting specific parameters of the electrocardiograph cycle
    • A61B5/352—Detecting R peaks, e.g. for synchronising diagnostic apparatus; Estimating R-R interval
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/02—Stethoscopes
    • A61B7/04—Electric stethoscopes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
    • A61B5/024—Detecting, measuring or recording pulse rate or heart rate
    • A61B5/0245—Detecting, measuring or recording pulse rate or heart rate by using sensing means generating electric signals, i.e. ECG signals

Definitions

  • FIG. 1 illustrates a block diagram of a coronary-artery-disease detection system
  • FIG.2 illustrates a first aspect of a data recording module and a first aspect of an associated docking system, in accordance with a first aspect of the coronary-artery-disease detection system illustrated in FIG. 1;
  • FIG. 3 illustrates a fragmentary view of a human thorax and associated prospective locations of auscultatory sound sensors at associated right, sternum and left, second, third, fourth and fifth, inter-costal spaces, left posterior locations at the second and third inter-costal spaces, and locations proximate to the heart apex;
  • FIG. 4 illustrates a second aspect of a data recording module and a second aspect of an associated docking system, in accordance with a second aspect of the coronary-artery-disease detection system illustrated in FIG. 1;
  • FIG. 5a illustrates an auscultatory sound sensor coupled to the skin of a test-subject, by bonding via associated adhesive layers or surfaces on both sides of an adhesive interface;
  • FIGS.5b and 5c each illustrate an auscultatory sound sensor that is detached, and therefore fully decoupled, from the skin of a test-subject, wherein FIG.5b illustrates the associated adhesive interface detached from the skin of the test-subject, and FIG. 5c illustrates the associated adhesive interface detached from the auscultatory sound sensor;
  • FIGS. 5d through 5g each illustrate an auscultatory sound sensor that is partially coupled to, but debonded from, the skin of a test-subject;
  • FIG. 6 illustrates a test-subject reclined on a surface, with their torso inclined while capturing auscultatory sound signals from a plurality of auscultatory sound sensors attached to the thorax of the test-subject;
  • FIG. 7 illustrates a flowchart of a first aspect of an associated auscultatory-sound-sensing process that incorporates a process for detecting a decoupling of the associated auscultatory sound sensors from the skin of the thorax of a test-subject being diagnosed for a prospective abnormal cardiovascular condition, wherein the decoupling-detection process occurs after each block of breath-held auscultatory sound time-series data is acquired, and is based upon scaled time-series data and responsive to an associated pre-determined debond-detection threshold;
  • FIG. 8 illustrates a flowchart of a first aspect of a process for acquiring auscultatory sound signals from the associated auscultatory sound sensors coupled to the skin of the thorax of the test- subject being diagnosed for a prospective abnormal cardiovascular condition;
  • FIG. 9 illustrates a plurality of six blocks of breath-held, auscultatory-sound-sensor time- series data recorded from an auscultatory sound sensor coupled to the skin of the thorax of a test- subject being diagnosed for a prospective abnormal cardiovascular condition;
  • FIGS. 10a- lOf respectively illustrate a simulation of successively recorded blocks of breath-held, sensor time-series data illustrated in FIG. 9, each illustrated with an expanded time scale, wherein FIGS. 10a- lOe illustrates a condition for which the auscultatory sound sensor is coupled to the skin of the test-subject, and FIG. lOf illustrates a condition for which the auscultatory sound sensor is decoupled from the skin of the test-subject;
  • FIG. 11 illustrates a flowchart of a process for determining a scale factor used to scale auscultatory-sound-sensor time-series data, the latter of which is analyzed to detect whether or not the associated auscultatory sound sensor is decoupled from the skin of the test-subject, wherein the scale factor provides for directly determining if the associated auscultatory sound sensor is detached from the skin of the test-subject;
  • FIGS. 12a-12f respectively illustrate time-series of the absolute values of the corresponding time-series data illustrated in FIGS. 10a- lOf, further illustrating a division of the block of breath-held, sensor time-series data into a plurality of associated data segments, with each data segment of sufficient width to nominally include sound from a single heartbeat, and with the peak values in each data segment marked, wherein FIGS. 12a-12e illustrates a condition for which the auscultatory sound sensor is coupled to the skin of the test-subject, and FIG. 12f illustrates a condition for which the auscultatory sound sensor is decoupled from the skin of the test-subject;
  • FIG. 13 illustrates an accelerometer on the thorax of a test-subject during a respiration cycle of the test-subject
  • FIG. 14 illustrates a breath-hold detection process
  • FIG. 15 illustrates a flowchart of a first aspect of a process for detecting whether or not an auscultatory sound sensor is debonded from the skin of a test-subject
  • FIG. 16 illustrates an organization of data from an auscultatory sound sensor recorded by an auscultatory coronary-artery-disease detection system from a test subject
  • FIG. 17 illustrates a flowchart of a noise detection process
  • FIG. 18 illustrates a flowchart of a process for generating a matched noise filter
  • FIG. 19 illustrates a flowchart of a process for evaluating the noise content in a spectral signal of an auscultatory sound signal
  • FIG. 20 illustrates a flowchart of a process for logging results from the noise-evaluation process of FIG. 19;
  • FIG. 21 illustrates a block diagram of a process for preprocessing and screening auscultatory sound signals
  • FIG. 22a illustrates a process for pre-processing auscultatory sound signals from auscultatory sound-or-vibration sensors
  • FIG. 22b illustrates a process for pre-processing electrographic signals from an ECG sensor
  • FIG. 23 illustrates a process for segmenting auscultatory sound signals from auscultatory sound-or-vibration sensors, by heart cycle based upon an electrographic signal from an ECG sensor, and by heat phase based upon the auscultatory sound signals;
  • FIG. 24 illustrates an auscultatory sound signal from an auscultatory sound-or-vibration sensor
  • FIG. 25 illustrates a corresponding electrographic signal from an ECG sensor, in correspondence with the auscultatory sound signal illustrated in FIG. 24;
  • FIG. 26 illustrates a process for identifying heart-cycle boundaries in an electrographic signal from an ECG sensor
  • FIG. 27 illustrates a first process for generating an envelope of a signal, which is called from the process illustrated in FIG. 26;
  • FIG. 28 illustrates an envelope, and associated peaks, of a portion of the electrographic signal illustrated in FIG. 25, generated in accordance with the process illustrated in FIGS. 26 and 27;
  • FIG. 29 illustrates a first aspect of a process for validating the peaks of an electrographic signal, which is called from the process illustrated in FIG. 26;
  • FIG. 30 illustrates a plot of the auscultatory sound signal illustrated in FIG. 24 together with plot of an envelope of the corresponding electrographic signal illustrated in FIGS. 25 and 28;
  • FIG. 31 illustrates a plot of the auscultatory sound signal illustrated in FIG. 24, but over a relatively greater total period of time, with the auscultatory sound signal segmented based upon the peaks of the envelope of the associated electrographic signal, and presented as a beat stack, each illustrating a corresponding associated heart beat;
  • FIG. 32 illustrates a second process for generating an envelope of a signal, which is called from the process illustrated in FIG. 23;
  • FIG. 33 illustrates a rectified auscultatory sound signal and an associate envelope thereof determined in accordance with the process illustrated in FIG. 32;
  • FIG. 34 illustrates a plot of the auscultatory sound signal that is shown rectified in FIG. 33, together with plots of the associated quadratic models of the associated envelopes in proximity to the associated SI and S2 heart sounds, illustrating time-points associated with zero-amplitude roots of the associated quadratic models that are used to locate associated heart phases;
  • FIG. 35 illustrates the auscultatory sound signal beat stack illustrated in FIG. 31, together with indicia showing the locations of the roots of the quadratic models associated with the SI and S2 heart sounds that are used to locate associated heart phases, with the auscultatory sound signals of the beat stack aligned with one another based upon the mean values of the roots of the S2 heart sounds;
  • FIG. 36 illustrates a process for identifying outliers in a diastole region of an auscultatory sound signal
  • FIG. 37 illustrates a flow chart of a process for selecting valid heart cycles and analyzing the results therefrom
  • FIG. 38 illustrates a portion an auscultatory sound signal during diastole, and also illustrates associated noise power thresholds
  • FIG. 39a illustrates an auscultatory sound signal for a plurality of heart cycles
  • FIG.39b illustrates the auscultatory sound signal for the plurality of heart cycles illustrated in FIG. 39a, with the heart cycles aligned with respect to the mean times of the S2 heart sound;
  • FIG. 40a illustrates the auscultatory sound signal for a plurality of heart cycles illustrated in FIG. 39b, including indications of the start of the S2 heart sound and the end of diastole;
  • FIG. 40b illustrates portions of the auscultatory sound signal during diastole for the plurality of heart cycles illustrated in FIG. 40a, with the heart cycles temporally normalized and resampled;
  • FIG. 41 illustrates a cross-correlation process
  • FIG. 42 illustrates a image of a plurality of localized cross-correlation signals associated with auscultatory sound signals from a plurality of heart cycles
  • FIG. 43 illustrates a time-frequency analysis based on a continuous wavelet transform of an auscultatory sound signal during diastole
  • FIG. 44 illustrates a process for detecting coronary artery disease from features of cross correlation images generated from analysis of auscultatory sound signals during diastole for a plurality of heart cycles, incorporating a Convolution Neural Network (CNN) classifier with a single convolution layer;
  • CNN Convolution Neural Network
  • FIG. 45 illustrates a display of results in a heart/arteries view and textual view
  • FIG. 46a illustrates a displace of a Receiver Operating Characteristic (ROC) curve
  • FIG.46b illustrates line graph display of true positives, true negatives, false negatives, and false positives
  • FIG.47 illustrates a display of a Heartbeat View comprising a graphical plot of the systolic and diastolic intervals of each heartbeat captured
  • FIG. 48 illustrated a display of a Stacked Heartbeat View
  • FIG. 49 illustrates a Bruit Identification View in a Line Graphs with Underfill mode
  • FIG. 50 illustrates a display of a Bruit Identification View in a Spectrogram mode
  • FIG. 51 illustrates a display of a Bruit Analysis View
  • FIG. 52 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Spectrogram mode, to confirm success of PCI;
  • FIG. 53 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Line Graph with Underfill mode, to confirm success of PCI;
  • FIG. 54 illustrates a second aspect of a process for detecting R-peaks in an electrographic signal
  • FIG. 55 illustrates an envelope, and associated peaks, of a portion of the electrographic signal that includes noise and other signal components to be ignored;
  • FIG.56 illustrates a second aspect of a process for validating the peaks of an electrographic signal, which is called from the process illustrated in FIG. 26, following the second aspect process for detecting R-peaks in an electrographic signal, illustrated in FIG. 26;
  • FIG. 57 illustrates a three-level Discrete Wavelet Transformation (DWT) process
  • FIG. 58 illustrates the effective filter spectra associated with the outputs of the three-level Discrete Wavelet Transformation process illustrated in FIG. 57;
  • FIG. 59 illustrates a scaling function of a Daubechies 4 (db4) wavelet family
  • FIG. 60 illustrates a wavelet function of a Daubechies 4 (db4) wavelet family
  • FIG. 61 illustrates a kernel discrete wavelet transformation process
  • FIG. 62 illustrates a three-level Wavelet Packet Transformation (WPT) process
  • FIG. 63 illustrates the effective filter spectra associated with the outputs of the three-level Discrete Wavelet Transformation process illustrated in FIG. 62;
  • FIG. 64 illustrates a high-pass filtered auscultatory sound signal for a single heart cycle from a high-pass filter having a 30 Hz cut-off frequency
  • FIG. 65 illustrates a Wavelet Packet Transformation energy map for the high-pass filtered auscultatory sound signal illustrated in FIG. 64, with either decomposition levels;
  • FIG. 66 illustrates a selected basis of frequency bins from the Wavelet Packet Transformation energy map illustrated in FIG. 64, selected using Shannon entropy as a cost function
  • FIG. 67 illustrates a Support Vector Machine classifier trained to distinguish two-class data samples with respect to an associated highest-margin hyperplane
  • FIG. 68 illustrates a feed-forward artificial neural network for binary classification
  • FIG. 69 illustrates a geometric construction that provides for determining a projected transverse location of an acoustic source with respect to a plane containing three auscultatory sound sensors
  • FIG. 70 illustrates various phases of a heart cycle in relation to an electrographic signal
  • FIG. 71 illustrates the SI, S2, S3 and S4 phases of a heart cycle in relation to an electrographic signal
  • FIG. 72 illustrates a second embodiment of a Convolution Neural Network (CNN).
  • CNN Convolution Neural Network
  • an auscultatory coronary-artery-disease detection system 10 incorporates at least one auscultatory sound sensor 12 that is operatively coupled to a recording module 13 running at least a first portion 14.1 of a Data Recording Application (DRA) 14, 14.1 on a first specialized computer or electronic system comprising a first computer processor or FPGA (Field Programmable Gate Array) 15 and first memory 17 powered by a first power supply 19, which provides for recording and preprocessing auscultatory sound signals 16 from the at least one auscultatory sound sensor 12.
  • DRA Data Recording Application
  • the at least one auscultatory sound sensor 12 comprises a first group 12’ of three auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ physically interconnected end-to-end with one another, and physically and electrically interconnected with a first wireless transceiver 18; and a second group 12” of three auscultatory sound sensors 12, 12 1 ”, 12 2 ”, 12 3 ” physically interconnected end-to-end with one another, and physically and electrically interconnected with the first wireless transceiver 18, with both groups 12% 12” of auscultatory sound sensors placed on the skin of the thorax 20 of a test-subject 22, in acoustic communication therewith.
  • the placement of the first group of auscultatory sound sensors 12’ in FIG. 1 is illustrated with the respective associated auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ in substantial alignment with the corresponding respective third R3, fourth R4 and fifth R5, inter-costal spaces on the right side 20 R of the thorax 20, and the placement of the second group of auscultatory sound sensors 12” in FIG. 1 is illustrated with the respective associated auscultatory sound sensors 12, 12 1 ”, 12 2 ”, 12 3 ” in substantial alignment with the corresponding respective third L3, fourth L4 and fifth L5, inter-costal spaces on the left side 20 L of the thorax 20.
  • prospective left-side posterior sensor locations LP2 and LP3 illustrated in FIG. 3 respectively refer to at the second LP2 and third LP3 intercostal spaces locations on the posterior of the thorax 20.
  • prospective sensor locations HA-1 and HA-2 are proximate to the apex of the heart, either on the anterior or the posterior of the thorax 20.
  • the auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ”, 12 1 ”, 12 2 ”, 12 3 ” are located at the second S2, L2, third S3, L3 and fourth S4, L4 inter-costal spaces at the sternum S2-S4 and left-side L2-L4 of the thorax 20.
  • the term“auscultatory sound” is intended to mean a sound originating from inside a human or animal organism as a result of the biological functioning thereof, for example, as might be generated by action of the heart, lungs, other organs, or the associated vascular system; and is not intended to be limited to a particular range of frequencies— for example, not limited to a range of frequencies or sound intensities that would be audible to a human ear, - but could include frequencies above, below, and in the audible range, and sound intensities that are too faint to be audible to a human ear.
  • the term“auscultatory-sound sensor” is intended to mean a sound sensor that provides for transducing auscultatory sounds into a corresponding electrical or optical signal that can be subsequently processed.
  • the auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” provide for transducing the associated sounds received thereby into corresponding auscultatory sound signals 16 that are preprocessed and recorded by an associated hardware-based signal conditioning/preprocessing and recording subsystem 25, then communicated to the first wireless transceiver 18, and then wirelessly transmitted thereby to an associated second wireless transceiver 26 of an associated wireless interface 26’ of an associated docking system 27, possibly running a second portion 14.2 of the Data Recording Application (DRA) 14, 14.2 on a corresponding second specialized computer or electronic system comprising an associated second computer processor or FPGA (Field Programmable Gate Array) 28 and second memory 30, both of which are powered by an associated second power supply 32, which together provide for recording and preprocessing the associated auscultatory sound signals 16 from the auscultatory sound sensors 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ”.
  • DDA
  • the hardware-based signal conditioning/preprocessing and recording subsystem 25 includes an amplifier— either of fixed or programmable gain,— a filter and an analog-to-digital converter (ADC).
  • the analog-to-digital converter (ADC) is a 16- bit analog-to-digital converter (ADC) that converts a -2.25 to +2.25 volt input to a corresponding digital value of - 32, 768 to + 32, 767.
  • the amplifier gain is programmable to one of sixteen different levels respectively identified as levels 0 to 15, with corresponding, respective gain values of 88, 249, 411, 571, 733, 894, 1055, 1216, 1382, 1543, 1705, 1865, 2027, 2188, 2350 and 2510, respectively for one set of embodiments.
  • the amplifier gain is fixed at the lowest above value, i.e., for this example, 88. so as to provide for avoiding the relative degradation of the associated signal-to-noise ratio (SNR) that naturally occurs with the relatively high gain levels of the programmable-gain set of embodiments.
  • SNR signal-to-noise ratio
  • DRA Data Recording Application
  • either or both the recording module 13 or docking system 27 may be constructed and operated in accordance with the disclosure of U.S. Provisional Application No. 62/575,364 filed on 20 October 2017, entitled CORONARY ARTERY DISEASE DETECTION SYSTEM, or International Application No. PCT/US2018/056832 filed on 22 October 2018, entitled CORONARY ARTERY DISEASE DETECTION SIGNAL PROCESSING SYSTEM AND METHOD, each of which is incorporated by reference in its entirety.
  • the auscultatory coronary-artery-disease detection system 10 may further incorporate an ECG sensor 34, for example, in one set of embodiments, an ECG sensor 34’ comprising a pair of electrodes incorporated in a corresponding pair of auscultatory sound sensors 12, wherein the signal from the ECG sensor 34’ is also preprocessed and recorded by a corresponding different signal channel of the same hardware-based signal conditioning/preprocessing and recording subsystem 25 of the recording module 13 that is used to preprocess the signals from the one or more auscultatory sound sensors 12.
  • an ECG sensor 34 comprising a pair of electrodes incorporated in a corresponding pair of auscultatory sound sensors 12, wherein the signal from the ECG sensor 34’ is also preprocessed and recorded by a corresponding different signal channel of the same hardware-based signal conditioning/preprocessing and recording subsystem 25 of the recording module 13 that is used to preprocess the signals from the one or more auscultatory sound sensors 12.
  • the ECG sensor 34 may comprise a separate set of a pair or plurality of electrodes that are coupled to the skin of the test subject, for example, in one set of embodiments, a pair of signal electrodes 35, 35 +/ in cooperation with a ground electrode 35°, wherein, referring to FIG. 3 (illustrating the locations of the electrodes 35, 35 +/ , 35°), in accordance with one non-limiting embodiment, the signal electrodes 35, 35 +/ span the heart from diametrically-opposed quadrants of the torso 44, and the ground electrode 35° is located in a different quadrant, orthogonally displaced from a midpoint of a baseline connecting the signal electrodes 35, 35 +/ .
  • the signal electrodes 35, 35 +/ could span across the top of the top of the thorax 20, or at any pair of locations commonly used for conventional ECG tests.
  • the recording module 13 and docking system 27 may each incorporate a corresponding respective USB interface 36.1, 36.2 to provide for transferring corresponding auscultatory sound signals 16 and or an electrographic signal 37 - from an associated ECG sensor 34, 34’— from the recording module 13 to the docking system 27, for example, rather than relying upon the first 18 and second 26 wireless transceivers of an associated wireless interface 26’.
  • data may be transferred from the recording module 13 to the docking system 27 via a portable memory element, for example, either an SD memory card or a Micro SD memory card.
  • a portable memory element for example, either an SD memory card or a Micro SD memory card.
  • the functionality of the Data Recording Application (DRA) 14 is distributed across the recording module 13 and the docking system 27.
  • the Data Recording Application (DRA) 14 spans across the recording module 13 and the docking system 27, with a first portion 14.1 comprising the hardware-based signal conditioning/preprocessing and recording subsystem 25 operative on the recording module 13, and a remaining second portion 14.2 operative on the docking system 27.
  • the Data Recording Application (DRA) 14 is operative entirely on the recording module 13.
  • the auscultatory sound sensor 12 provides for sensing sound signals that emanate from within the thorax 20 of the test-subject 22 responsive to the operation of the test-subject’s heart, and the resulting flow of blood through the arteries and veins, wherein an associated build-up of deposits therewithin can cause turbulence in the associated blood flow that can generate associated cardiovascular-condition-specific sounds, the latter of which can be detected by a sufficiently- sensitive auscultatory sound sensor 12 that is acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22.
  • the sound level of these cardiovascular-condition-specific sounds can be below a level that would be detectable by a human using a conventional stethoscope.
  • these sound levels are susceptible to detection by sufficiently sensitive auscultatory sound sensor 12 that is sufficiently acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22.
  • the auscultatory sound sensor 12 may be constructed in accordance with the teachings of U.S. Provisional Application No. 62/568,155 filed on 04 October 2017, entitled AUSCULTATORY SOUND SENSOR, or International Application No.
  • the auscultatory sound sensor 12 may be constructed in accordance with the teachings of U.S. Pat. Nos. 6,050,950, 6,053,872 or 6,179,783, which are incorporated herein by reference. Referring also to FIG.
  • the auscultatory sound sensors 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” are acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22 via an acoustic interface 40, for example, via a hydrogel layer 40’, that also functions as an associated adhesive interface 42 that is attached to the associated auscultatory sound sensor 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” with a first adhesive layer 42.1, for example, either a first surface 40.1’ of the hydrogel layer 40’ or a first layer of double-sided tape 42.1’ on a first side of the acoustic/adhesive interface 40, 42, and that is attached to the skin 38 of the thorax 20 of the test- subject 22 with a second adhesive layer 42.2, for example, either a second surface 40.2’ of the hydrogel layer 40’ or a second layer of double-sided
  • the auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 is fully attached to the acoustic/adhesive interface 40, 42 via the first adhesive layer 42.1, 42.1% 40.1% and the acoustic/adhesive interface 40, 42 is fully attached to the skin 38 of the thorax 20 of the test- subject 22 via the second adhesive layer 42.2, 42.2% 40.2% so that sound signals from within the thorax 20 of the test-subject 22 can propagate otherwise unimpeded to the auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12' " , 12 2 ”, 12 3 ”, thereby providing for a maximum achievable level of the corresponding associated auscultatory sound signals 16, and thereby improving the prospect of detecting an associated abnormal cardiovascular condition - if present - therefrom.
  • the auscultatory sound sensor 12, 12 r , 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 is partially attached to the skin 38 of the thorax 20 of the test-subject 22, and thereby partially decoupled therefrom - i.e., in a condition referred to herein as being debonded therefrom— the resulting auscultatory sound sensor 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” would be only partially responsive to sound signals from within the thorax 20 of the test-subject 22, but not sufficiently responsive to provide for an associated auscultatory sound signal 16 of sufficient amplitude to provide for reliably detecting a prospective associated abnormal cardiovascular condition.
  • FIGS. 5d and 5e respectively illustrate an acoustic/adhesive interface 40, 42 partially detached from skin 38, and an acoustic/adhesive interface 40, 42 partially detached from an auscultatory sound sensor 12, respectively.
  • FIG. 5f illustrates an auscultatory sound sensor 12 attached to a wrinkled acoustic/adhesive interface 40, 42
  • FIG. 5g illustrates an acoustic/adhesive interface 40, 42 attached to wrinkled skin 38.
  • DAA Data Recording Application
  • auscultatory sound sensors 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 are, or are, either detached or debonded from the skin 38 of the thorax 20 of the test-subject 22, so as to provide for excluding data from auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” that are either detached or debonded, from the skin 38 of the thorax 20 of the test-subject 22 from being used to diagnose a prospective abnormal cardiovascular condition.
  • the adhesive interface 42 could comprise either a hydrogel layer 40’, for example, P-DERM® Hydrogel; a silicone material, for example, a P-DERM® Silicone Gel Adhesive; an acrylic material, for example, a P-DERM® Acrylic Adhesive; a rubber material; a synthetic rubber material; a hydrocolloid material; or a double-sided tape, for example, with either rubber, acrylic or silicone adhesives.
  • the quality of the auscultatory sounds acquired from a test-subject 34 can be improved if the torso 44 of the test-subject 34 is inclined, for example, in one set of embodiments, at an inclination angle Q of about 30 degrees above horizontal— but generally, as close to upright (i.e.
  • an auscultatory-soundsensing process 700 provides for first determining a scale factor SF from an initially-acquired block of auscultatory-sound-sensor time-series data S, and initially determining from the scale factor SF whether one or more of the auscultatory sound sensors 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” is detached from the skin 38 of the thorax 20 of the test-subject 22, wherein when multiplied by the scale factor SF, the values of the associated auscultatory-sound-sensor time-series data S are nominally within a range that is a predetermined percentage of the dynamic range of the associated data acquisition system (for example, that provides 16- bit signed digital values).
  • the auscultatory-sound-sensing process 700 provides for acquiring successive blocks of auscultatory-sound-sensor time-series data S while the test-subject 22 is holding their breath, and determining from each block of auscultatory- sound-sensor time-series data S - using an associated predetermined debond-detection threshold
  • auscultatory-sensor time-series data S is rejected if excessive noise is detected, and the test is aborted if one or more auscultatory sound sensors 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” has become decoupled from the skin 38 of the thorax 20.
  • the first aspect 700 of the auscultatory-soundsensing process 700 commences with step (702) by initializing a data-block counter i to a value of zero, and then, in step (704), acquiring a block of Ns contiguous samples of auscultatory- sound-sensor time-series data S in accordance with a first aspect 800 of a data acquisition process 800.
  • This initially-acquired data is then used to determine the scale factor SF that is used to determine whether or not one or more auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” is/are detached from the skin 38 of the thorax 20 of the test-subject 22, and then subsequently to scale subsequent blocks of time-series data S.
  • the initial block of auscultatory- sound-sensor time-series data S may be acquired either with, or without, the test-subject 22 holding their breath, but typically with the test-subject 22 allowed to breath normally— for their comfort and convenience.
  • the number of samples Ns to be acquired is given by the product of an acquisition period Si in seconds, times a sampling frequency Fs in Hz.
  • the data acquisition process 800 commences with step (802) by pre-filtering the electronic signal from the associated auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” with an analog anti-aliasing filter, for example, an analog second-order band-pass filter having a pass band in the range of 3 Hz to 2.5 KHz, for which the upper-cutoff frequency is sufficiently below the sampling frequency (i.e. no more than half) so as to prevent high frequency components in the signal being sampled from appearing as low frequency components of the sampled signal, i.e. so as to prevent aliasing.
  • an analog anti-aliasing filter for example, an analog second-order band-pass filter having a pass band in the range of 3 Hz to 2.5 KHz, for which the upper-cutoff frequency is sufficiently below the sampling frequency (i.e. no more than half) so as to prevent high frequency components in the signal being sampled from appearing as low frequency components of the sampled signal, i.e. so
  • step (804) the test-subject 22 need not necessarily hold their breath - as is the case for the initially-acquired block of auscultatory-sound-sensor time-series data S,— then, in step (806), the pre-filtered auscultatory sound signal 16 is sampled at the sampling frequency Fs and converted to digital form by the associated analog-to-digital (ADC) converter.
  • ADC analog-to-digital
  • FIGS. 9 and 10a each illustrate a first block of auscultatory-sound-sensor time-series data S of duration So that was recorded from one of the auscultatory sound sensors 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ”.
  • the scale factor SF is determined from the initially-acquired block of auscultatory-sound-sensor time-series data S, in accordance with an associated scale-factor-determination process 1100. More particularly, referring to FIGS. 11 and 12a, the scale-factor-determination process 1100 commences with step (1102), wherein the block of auscultatory-sound-sensor time-series data S is divided into a plurality of data segments 46, for example, each of the same data-segment duration SD that nominally spans a single heartbeat, for example, about one second. For example, in FIG.
  • FIG. 12a illustrates a time series
  • X the maximum absolute value of the auscultatory-sound-sensor time-series data S(k) is determined, as given by:
  • step (1104) the median of these maximum values is determined, as given by
  • step (1106) the scale factor SF is determined, as given by:
  • DRo-p is the nominal zero-to-peak dynamic range of the auscultatory-sound-sensor time-series data S after scaling, i.e. after multiplying the acquired values by the scale factor SF.
  • the nominal zero-to-peak dynamic range is set to be about 80 percent - more broadly, but not limiting, 80 percent plus or minus 15 percent— of the zero-to- peak range of the associated analog-to-digital converter - for example, in one set of embodiments, a 16-bit signed analog-to-digital converter— used to digitize the auscultatory-sound-sensor time- series data S in step (806).
  • the scale factor SF is integer-valued that, for an attached and bonded auscultatory sound sensor 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ”, ranges in value between 1 and 28.
  • the associated level of the auscultatory sound signals 16 will be low - for example, at a noise level - resulting in a relatively large associated scale factor SF from step (1106).
  • step (1108) the scale factor SF is in excess of an associated threshold SFMAX
  • the Data Recording Application (DRA) 14 is aborted in step (1110), and the operator 48 is alerted that the one or more auscultatory sound sensors 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” is/are detached, so that this can be remedied.
  • the value of the threshold SFMAX is 28 for the above-described fixed-gain embodiment, i.e.
  • step (1108) if the value of the scale factor SF does not exceed the associated threshold SFMAX, in step (1112), the scale factor SF is returned to step (706) for use in scaling subsequently -recorded breath-held auscultatory sound signals 16.1
  • step (708) the value of the data-block counter i is incremented, so as to point to the next block of auscultatory-sound-sensor time- series data S to be recorded. If, while this next block is recorded, the auscultatory sound sensors 12, 12 1 ’, 12 , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” remain attached and bonded to the skin 38 of the thorax 20 of the test-subject 22, and the associated breath-held auscultatory sound signals 16.1 are not corrupted by excessive noise, then this next block of auscultatory-sound-sensor time-series data S will then be subsequently used to detect a prospective abnormal cardiovascular condition therefrom.
  • the auscultatory sound signals 16 used to detect prospective abnormal cardiovascular conditions are recorded while the test-subject 22 holds their breath, the latter to prevent the resulting cardiovascular-based auscultatory sound signals 16 from being overwhelmed by breathing-related sounds that are substantially louder than cardiovascular-based sounds, thereby providing for improving the associated signal-to-noise ratio of the cardiovascular- based auscultatory sound signals 16.
  • a next block of Ns contiguous samples of auscultatory-soundsensor time-series data S is acquired over an acquisition period Si in accordance with a first aspect 800 of a data acquisition process 800, during which time the test-subject 22 is instructed to hold their breath.
  • the data acquisition process 800 commences with step (802) by pre-filtering the electronic signal from the associated auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 , 12 1 ”, 12 2 ”, 12 3 ” with the above-described analog anti-aliasing filter. Then, from step (804), because breath-held data is to be acquired, in step (812), the test-subject 22 is instructed by the operator 48 to hold their breath.
  • ADC analog-to-digital
  • step (820) if one or more addition samples remain to be acquired, and if the operator 48 continues to observe that the test-subject 22 is holding their breath, or, additionally or alternatively, if this is confirmed by a below-described breath-hold detection process 1400, then, in step (822) the sample counter j is incremented, and the next sample is acquired in step (818).
  • the pre-filtered auscultatory sound signals 16 are also separately-recorded while waiting for the test-subject 22 to hold their breath, or resume doing so.
  • the auscultatory sound signals 16 typically continue to be recorded between breath-held segments, that latter of which are identified by associated recorded start and stop times with respect to the associated continuous recording.
  • the auscultatory coronary-artery-disease detection system 10 may further incorporate an accelerometer 50 operatively coupled to the thorax 20 of the test-subject 22 to provide an associated acceleration signal responsive to the motion of the thorax 20.
  • an accelerometer 50 operatively coupled to the thorax 20 of the test-subject 22 to provide an associated acceleration signal responsive to the motion of the thorax 20.
  • the associated acceleration signal - operatively coupled to recording module 13 and possibly transmitted to the docking system 27— may be twice integrated either in recording module 13 or the docking system 27 to generate a measure of the peak-to-peak displacement of the thorax 20, which if greater than a threshold would be indicative of breathing by the test- subject 22. More particularly, referring also to FIG.
  • an acceleration signal 52 therefrom may alternatively or additionally be processed by an associated breath-hold detection process 1400 to provide for automatically determining - for example, in step (814) of the data acquisition process 800 illustrated in FIG. 8 - whether or not the test-subject 22 is breathing, responsive to the determination, from the acceleration signal 52, of a peak-to-peak displacement of the thorax 20 of the test-subject 22. More particularly, beginning with, and in, step (1402), the respective previous/initial values of thorax displacement Yo and thorax velocity Vo.
  • a sample counter i is initialized to an initial value, for example, zero; the respective minimum YMIN and maximum YMAX values of thorax displacement are each set equal to the (initial) value of thorax displacement Yo ; the values of the sample counter IMIN and IMAX at which the corresponding minimum YMIN and maximum YMAX values of thorax displacement occur are set equal to the initial value of the sample counter i; and a BreathingFlag is set to indicate that the test-subject 22 is breathing.
  • step (1404) the current sample of thorax acceleration A is acquired.
  • step (1406) if the sample counter i is equal to the initial value, i.e.
  • step (1412) the current value of thorax velocity V is calculated by integrating the previous Ao and current A measurements of thorax acceleration, for example, using a trapezoidal rule, as follows:
  • step (1414) the current value of thorax displacement Y is calculated by integrating the above-calculated previous Vo and current V values of thorax velocity, for example, again using a trapezoidal rule, as follows:
  • step (1416) the respective previous values of thorax acceleration Ao, thorax displacement Yo and thorax velocity Vo are respectively set equal to the corresponding current values of thorax acceleration A, thorax velocity V and thorax displacement T, respectively, that will each be used in subsequent iterations of steps (1412) and (1414).
  • step (1418) if the current value of thorax displacement Y is greater than then current maximum value of thorax displacement TMAX - for example, as would result during a phase of chest expansion by the test-subject 22,— then, in step (1420), the current maximum value of thorax displacement YMAX is set equal to the current value of thorax displacement Y and the corresponding value of the sample counter IMAX associated therewith is set to the current value of the sample counter i.
  • step (1422) if, in step (1422), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMAX associated with the maximum value of thorax displacement YMAX is not equal to a threshold value A (the relevance of which is described more fully hereinbelow), then in step (1424), if the current value of thorax displacement Y is less than then current minimum value of thorax displacement YMIN - for example, as would result during a phase of chest contraction by the test- subject 22, - then, in step (1426), the current minimum value of thorax displacement YMIN is set equal to the current value of thorax displacement Y and the corresponding value of the sample counter IMIN associated therewith is set to the current value of the sample counter i.
  • a threshold value A the relevance of which is described more fully hereinbelow
  • step (1428) if the amount by which the current maximum value of thorax displacement YMAX is greater the current minimum value of thorax displacement YMIN meets or exceeds a displacement threshold A YMAX. then, in step (1430), the BreathingFlag is set to indicate that the test-subject 22 is breathing, after which, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404). Similarly, from step (1428), if the displacement threshold AYMAX is not exceeded, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404).
  • step (1424) for example, as would result from other than a phase of chest contraction by the test-subject 22,— if, in step (1432), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMIN associated with the minimum value of thorax displacement YMIN is not equal to the threshold value A, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404).
  • step (1432) If, from step (1432), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMIN associated with the minimum value of thorax displacement YMIN is equal to the threshold value A— following a minimum chest contraction of the test-subject 22, in anticipation of subsequent chest expansion, wherein the threshold value A is greater than or equal to one,— then, in step (1434), the peak-to-peak thorax displacement AY is calculated as the difference between the current maximum YMAX and minimum YMIN values of thorax displacement, and, in step (1436), the maximum value of thorax displacement YMAX is set equal to the current value of thorax displacement Y, and the value of the sample counter IMAX at which the corresponding maximum value of thorax displacement YMAX occurred is set equal to the current value of the sample counter i, in anticipation of subsequently increasing magnitudes of the current value of thorax displacement Y to be tracked in steps (1418) and (1420).
  • step (1422) the amount by which the current value of the sample counter i exceeds the value of the sample counter IMAX associated with the maximum value of thorax displacement YMAX is equal to the threshold value A— following a maximum chest expansion of the test-subject 22, in anticipation of subsequent chest contraction, wherein the threshold value A is greater than or equal to one,— then, in step (1438), the peak-to-peak thorax displacement AY is calculated as the difference between the current maximum YMAX and minimum YMIN values of thorax displacement, and, in step (1440), the minimum value of thorax displacement YMIN is set equal to the current value of thorax displacement Y, and the value of the sample counter IMIN at which the corresponding minimum value of thorax displacement YMIN occurred is set equal to the current value of the sample counter i, in anticipation of subsequently decreasing magnitudes of the current value of thorax displacement Y to be tracked in steps (1424) and (1426).
  • the threshold value A provides for a delay to assure that a most-recent extremum of displacement has been reached, either the current maximum YMAX or minimum YMIN values of thorax displacement, before calculating the associated peak-to-peak thorax displacement AY.
  • step (1442) if the amount of the peak-to-peak thorax displacement AY calculated in steps (1434) or (1438), respectively, meets or exceeds the displacement threshold AYMAX, then, in step (1444), the BreathingFlag is set to indicate that the test-subject 22 is breathing. Otherwise, from step (1442), if the amount of the peak-to-peak thorax displacement AY calculated in steps (1434) or (1438), respectively, does not exceed the displacement threshold AYMAX, then, in step (1446), the BreathingFlag is reset to indicate that the test-subject 22 is not breathing. Following either step (1444) or (1446), in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404).
  • a corresponding block of scaled auscultatory- sound-sensor time-series data S is calculated by multiplying the acquired block of auscultatory- sound-sensor time-series data S by the scale factor SF, and in step (714), the scaled auscultatory-sound-sensor time-series data S is analyzed by an associated debond detection process 1500 to determine whether or not any of the auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” was debonded from skin 38 of the thorax 20 of the test-subject 22 during the associated data acquisition process 800.
  • the debond detection process 1500 commences with step (1502) by initializing a sample counter j to a value of one, and initializing a threshold counter TC to a value of zero, wherein the threshold counter TC is a count of the number of contiguous successive samples for which the value of the scaled auscultatory-sound-sensor time-series data S is less than an associated predetermined debond-detection threshold DT.
  • the debond-detection threshold DT is set to a value that is about 20% of the achievable maximum value of the output from the analog-to-digital converter (ADC).
  • step (1504) if the absolute value of the sample of scaled auscultatory-soundsensor time-series data S, i.e.
  • step (1510) if the sample counter j does not exceed the number of samples Ns in the block of scaled auscultatory-sound-sensor time-series data S, then, in step (1512), the sample counter j is incremented, and the process continues again with step (1504). Otherwise, from step (1504), if the absolute value of the current sample of scaled auscultatory-sound-sensor time- series data S, i.e.
  • step (1510) if the sample counter j exceeds the number of samples Ns in the block of scaled auscultatory-sound-sensor time-series data S or auscultatory-sound-sensor time-series data S— indicating that the entire block of scaled auscultatory-sound-sensor time-series data S or auscultatory-sound-sensor time-series data S has been screened,— then the debond detection process 1500 is terminated with step (1516) by returning an indication to step (714) of the auscultatory-sound-sensing process 700 that the associated auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” is not debonded from the skin 38 of the thorax 20 of the test-subject 22.
  • step (1518) the debond detection process 1500 is terminated with step (1518) by returning an indication to step (714) of the auscultatory-soundsensing process 700 that the associated auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” is debonded from the skin 38 of the thorax 20 of the test-subject 22.
  • the value of N D is equal to 4, and the value of & is equal to 1 second.
  • step (716) if, from step (714), the debond detection process 1500 detected that the associated auscultatory sound sensor 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” was debonded while acquiring the block of auscultatory-sound-sensor time-series data S, then the Data Recording Application (DRA) 14 is aborted in step (718), and the operator 48 is alerted that one or more auscultatory sound sensors 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” are debonded, so that this can be remedied.
  • DRA Data Recording Application
  • step (720) if, from step (714), the debond detection process 1500 did not detect that the associated auscultatory sound sensor 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” was debonded while acquiring the block of auscultatory-sound-sensor time-series data S, then, in step (720), if sufficient noise-screened data has not been gathered - for example, in one set of embodiments, a total duration of at least 65 seconds of recorded data,— then the auscultatory- sound-sensing process 700 continues with step (708).
  • an associated noise detection (i.e. noise-screening) process - operating on either the block of scaled auscultatory-sound-sensor time-series data S, or the block of auscultatory-sensor time-series data S, in parallel with the debond detection process 1500 - provides for detecting if the block of auscultatory-sound-sensor time-series data S has been corrupted by excessive noise, and if so, from step (726), that block of auscultatory-sound-sensor time-series data S is ignored, and the auscultatory-sound-sensing process 700 continues by repeating step (710) to acquire a new block of auscultatory-sound-sensor time-series data S. Otherwise, from step (726), if the block auscultatory-sound-sensor time-series data S has not been corrupted by excessive noise, the process continues with the above-described step (720).
  • noise detection i.e. noise-screening
  • step (720) if sufficient noise-screened data has been gathered for which the associated one or more auscultatory sound sensors 12, 12 1 ’, 12 2 ’, 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” were not debonded from the skin 38 of the thorax 20 of the test-subject 22 - for example, in one set of embodiments, a total duration of at least 65 seconds of recorded data,— then, in step (722), at least the composite set of blocks of breath-held auscultatory-sound-sensor time-series data S acquired in step (710) are subsequently analyzed by an associated Data Analysis Application (DAA) 54 operative on the docking system 27 - as illustrated in FIGS.
  • DAA Data Analysis Application
  • FIGS. 9 and 10a- lOf illustrate a simulation of six blocks of breath-held auscultatory- sound-sensor time-series data S recorded in accordance with the first aspect 700 of auscultatory-sound-sensing process 700, with respective durations of ⁇ 3 ⁇ 4, ⁇ 3 ⁇ 4, ⁇ %, ⁇ 3 ⁇ 4, Ss, and ⁇ 3 ⁇ 4 during which time periods the test-subject 22 was holding their breath, separated by periods Di , 2h, , 2b, and 2b of normal breathing, wherein FIGS.
  • FIG. lOa-lOe illustrate breath-held auscultatory sound signals 16.1, 16.1’ from a normally-bonded auscultatory sound sensor 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ” as illustrated in FIG. 5a
  • FIG. lOf illustrates a breath-held auscultatory sound signal 16.1, 16.1” from a debonded auscultatory sound sensor 12, 12 1 ’, 12 r , 12 3 ’ , 12 1 ”, 12 2 ”, 12 3 ”, for example, as illustrated in any of FIGS.
  • FIGS. lOb-lOe are identical to FIG. 10a. However, it should be understood that typically the amplitude of the auscultatory sound signals 16, 16.1 varies from heartbeat to heartbeat, and from one breath-held segment to another.
  • one or more of the auscultatory-sound-sensing process 700, the data acquisition process 800, the scale-factor-determination process 1300, or the de-bond detection process 1500 could be implemented with corresponding alternative processes disclosed in U.S. Application Serial No. 16/136,015 filed on 19 September 2018 - with particular reference to FIGS. 16 through 22 - which is incorporated by reference herein in its entirety.
  • index pointer arrays io[] and h[] to store the sample locations at the beginning and end of breath-held data segments of the corresponding sampled auscultatory sound data S m [] from the m th auscultatory sound sensor 12, and later-used index pointer arrays isi[] and [] to store the sample locations of the SI sound at the beginning of each heart cycle, and the S2 sound at the beginning of diastole respectively, wherein in FIG.
  • a status array Status [m, k] indicates the measurement status of the k th breath-held data segment of the m th auscultatory sound signal 16, i.e. the sampled auscultatory sound data S m () from the m th auscultatory sound sensor 12. Accordingly, step (728)—that provides for ignoring data— may be implemented by setting the corresponding value of the status array Status(m, k) to a value of IGNORE. Referring to FIGS.
  • a noise detection process 1700 called from step (724) provides for determining whether or not a particular segment of breath-held auscultatory sound signal 16.1 is corrupted by excessive noise, and if so, provides for flagging that segment as being excessively noisy so as to be prospectively excluded from subsequent processing.
  • the noise detection process 1700 generates, in step (1714), frequency-domain noise filters FH[] responsive to cross-correlations of frequency spectra of pairs of adjacent auscultatory sound sensors 12. Accordingly, sensor-index pointers ml[p] and m2[p] provide for identifying the associated auscultatory sound sensors 12, ml[p], m2 [ p ] of each pair, the latter of which is identified by a pair pointer p.
  • the noise detection process 1700 For each pair of adjacent auscultatory sound sensors 12 in a set of adjacent pairs— selected by sensor-index pointers ml[p] and m2[p] in steps (1706) and (1710), respectively, wherein the pair pointer p is initialized to 1 in step (1704) - the noise detection process 1700 generates, in step (1714), a frequency-domain noise filter FH[] by cross- correlating, in step (1802), the frequency spectra FS A [] and FS B []— generated in steps (1708) and (1712)— of each associated breath-held auscultatory sound signal 16.1: S A [] and S B [], wherein the values of the frequency-domain noise filter FH[] are generated by normalizing the frequency spectra of the cross-correlation function to a range of 0 to 1 in step (1804), then subtracting these values from unity, and then setting resulting values that are less than a noise floor to the value of the noise floor in step (1806).
  • the frequency-domain noise filter FH[] provides for attenuating the components of the breath-held auscultatory sound signal 16.1: SA[] and SB[] that are correlated with one another.
  • step (1904) provides for a matched filtering process to accentuate the underlying noise in the associated breath-held auscultatory sound signal 16.1: SA[] or SB[] by attenuating frequency-component portions thereof that are correlated with a corresponding breath-held auscultatory sound signal 16.1: SB[] or SA[] from the other auscultatory sound sensor 12, ml[p], m21 p
  • step (1904) the product of the frequency-domain noise filter FH[] with either of the frequency spectra FS A [] or FS B [] is inverse Fourier transformed back to the corresponding time domain noise signal SN[]
  • STFT short-time Fourier transform
  • the average powers PLOW, PMID, PHIGH in three respective frequency ranges 20 Hz to 200 Hz, 200 Hz to 800 Hz, and 800 Hz to 1,500 Hz, respectively, is calculated in steps (1922), (1924) and (1926), respectively, wherein each average power PLOW, PMID, PHIGH is compared with a corresponding threshold— for example, in one set of embodiments, -20 dB, -50 dB and -30 dB, respectively - in step (1928).
  • the corresponding breath-held auscultatory sound signal 16.1: SA[] or SB[] is then flagged in step (1930) as being noisy if any of the associated noise power thresholds are exceeded.
  • one auscultatory sound sensor 12 exceeds the noise threshold in a given k th breath- held segment of breath-held sampled auscultatory sound data S m [io[k]: ii [k] ] , that auscultatory sound sensor 12, m is flagged in step (2008) as being noisy so as to later be ignored for that k th breath-held segment. If a particular auscultatory sound sensor 12, m exceeds the noise threshold for more than one breath-held segment of breath-held sampled auscultatory sound data S m [io[k]: ii[k]], then that auscultatory sound sensor 12, m is flagged in step (2016) as being bad.
  • the process of steps (1706) through (1722) repeats for each of NPAIRS pairs of adjacent auscultatory sound sensors.
  • the process of steps (1704) through (1726) is repeated until all the breath-held segments k of data have been processed.
  • the noise detection process 1700 commences with step (1702) by initializing a breath-held-segment pointer A to a value of 1, so as to provide for pointing to the first breath-held segment of data.
  • the breath-held-segment pointer k provides for pointing to the k th breath-held segment of data, of duration d A , extending between sample locations io[k] and ii[k], as illustrated in FIG. 16.
  • the pair pointer p is initialized to a value of 1
  • a noise counter NNOISY is initialed to a value of 0.
  • step (1710) the k th breath-held segment of breath-held sampled auscultatory sound data S m2
  • k]: ii[k]] is selected as the sampled auscultatory sound data S B [] of the second auscultatory sound sensor 12, ni2[p] of the pair p, and in step (1712) the Fourier Transform of the sampled auscultatory sound data S A [] is calculated as FS B [] FFT(S B []).
  • step (1714) the frequency-domain noise filter FH[] generated by a matched- noise-filter generation process 1800 that - referring also to FIG. 18 - commences in step (1802) with the cross correlation of the frequency spectra FS A [], FS B [] of the sampled auscultatory sound data S A [], S B [] of the first ml[p] and second m2[p] auscultatory sound sensors 12 of the pair p, wherein the resulting cross-correlation is stored in array FH[] and then normalized to a range of 0 to 1 in step (1804), and then inverted in step (1806), wherein each element of the normalized array FH[] is subtracted from 1, and if the result is less than a noise floor, is set to the value of the noise floor NoiseFloor.
  • step (1808) the resulting frequency-domain noise filter FH[] is returned and subsequently used in steps (1716) and (1718) of the noise detection process 1700 to evaluate the noise in the frequency spectra FS A [], FS B [] of the sampled auscultatory sound data S A [], S B [] of the first ml[p] and second m2[p] auscultatory sound sensors 12 of the pair p, respectively, in accordance with an associated noise-content- evaluation process 1900, the latter of which is called from step (1716) to evaluate the noise content of the frequency spectrum FS A [] of the sampled auscultatory sound data S A [] of the first auscultatory sound sensor 12, ml[p] of the pair p, and which is called from step (1718) to evaluate the noise content of the frequency spectrum FS B [] of the sampled auscultatory sound data S B [] of the second auscultatory sound sensor 12, m2[p] of the pair p.
  • the noise-content-evaluation process 1900 commences in step (1902) with receipt of the index m of the associated auscultatory sound sensor 12, m, the associated frequency spectrum FS[] of the associated auscultatory sound sensor 12, m, and the associated frequency-domain noise filter FH[] Then, in step (1904), the time domain noise signal SN[] - containing a total of NPTS data points, i.e.
  • the number of data points in the breath- held sampled auscultatory sound data S m [io[k]: ii[k]] - is given by the inverse Fourier Transform of the product of the associated frequency spectrum FS[] with the associated frequency-domain noise filter FH[], corresponding to a time-domain cross-correlation of the corresponding associated time-domain signals.
  • each of the NFFT summation data points of a frequency-domain summation array FXSUM[] is initialized to zero, as is an associated summation counter NSUM, wherein the number of summation data points NFFT is a power of 2, for example, 1024, and substantially less than the total number of data points NPTS in the time domain noise signal SN[]
  • an index j MiN — to the first sample of an N FFT -point window of samples to be analyzed from the time domain noise signal SN[] - is initialized to a value of 1.
  • step (1910) an index j ⁇ i ⁇ x— to the last sample of the N FFT - point window of samples to be analyzed from the time domain noise signal SN[] - is set to the value of jnii N + NFFT - 1. Then, in step (1912), if the end of the time domain noise signal SN[] has not been reached, then, in step (1914), the square of values - i.e.
  • step (1916) corresponding to noise power — of an N FFT -point Fourier Transform of the data from the N FFT -point window of samples of the time domain noise signal SN[] over the range of samples SN[JMIN] to SN[j MA x], is added to the frequency-domain summation array FXSUM[] Then, in step (1916), the summation counter NSUM is incremented, and in step (1918), the index j MiN is incremented by half the width of the N FFT -point window, i.e. by a value of NFFT 12, so as the provide for the next N FFT -point window to be analyzed to overlap the current N FFT -point window by half the window width.
  • the above noise-content-evaluation process 1900 repeats beginning with step (1910), until, in step (1912), the index j M A X exceeds the end of the time domain noise signal SN[], after which, in step (1920), each of the values of the frequency-domain summation array FXSUM[] is divided by the value of the summation counter NSUM SO as to calculate an associated average noise power FX[] Then, in steps (1922), (1924) and (1926), respectively, the noise power is summed in three different corresponding respective frequency ranges, for example, 20 Hz to 200 Hz, 200 Hz to 800 Hz, and 800 Hz to 1,500 Hz, respectively, to give three corresponding respective power levels, PLOW, PMID and PHIGH, respectively.
  • step (1928) if any of the power levels, PLOW, PMID or PHIGH exceeds a corresponding respective power threshold value ThresholdL O w, ThresholdMiD or Thresholdm GH , then, in step (1930), the noise threshold is considered to have been exceed for the particular auscultatory sound sensor 12, m and the particular associated segment k of breath-held sampled auscultatory sound data S m [io[k]: ii[k]] is flagged with a corresponding indication of an associated status in an associated NoiseStatus flag.
  • step (1932) the noise threshold is considered to have not been exceed for the particular auscultatory sound sensor 12, m and the particular associated segment of breath-held sampled auscultatory sound data S m [io[k]: ii [k] ] is flagged with a corresponding indication of the associated status in the NoiseStatus flag.
  • the results from either steps (1930) or (1932) are then logged by an associated results-logging process 2000, which is called from step (1934).
  • the results-logging process 2000 commences with receipt in step (2002) of the index m of the associated auscultatory sound sensor 12, m and the associated NoiseStatus flag. If, in step (2004), the NoiseStatus flag indicates a noisy auscultatory sound sensor 12, m, then, in step (2006), the noise counter NNOISY is incremented, and, in step (2008), the corresponding element of the status array Status [m, k] for the associated auscultatory sound sensor 12, m and breath-held segment k is updated to activate an associate noisy flag, so as to indicate the associated auscultatory sound sensor 12, m being noisy for the associated breath- held segment k.
  • step (2010) if the value of the noise counter NNOISY is greater than 1, then, in step (2012), the corresponding elements of the status array Status [m, k] for each of the auscultatory sound sensors 12, m is updated to activate the Ignore flag, so as to provide for ignoring each of the auscultatory sound sensors 12, m for the associated breath-held segment k.
  • step (2014) if the noisy flag of the status array Status [m, k] is activated for at least NFAIL breath-held segments k for the associated auscultatory sound sensor 12, m, then, in step (2016), status array Status [m] for the associated auscultatory sound sensor 12, m is updated to activate the Bad flag for the associated auscultatory sound sensor 12, m, so as to indicate that the associated auscultatory sound sensor 12, m is bad. Then, or otherwise from either step (2004) or step (2014), the results-logging process 2000 terminates by returning to the noise detection process 1700.
  • step (1720) if all of the pairs of adjacent auscultatory sound sensors 12 have not been processed, then, in step (1722), the pair pointer p is incremented, and the noise detection process 1700 is repeated for the next pair of auscultatory sound sensors 12, ml[p], ni2[p] beginning with step (1706) for the same breath-held segment k.
  • step (1724) If in step (1724), additional segments of breath-held sampled auscultatory sound data S m [io[k] : ii [k] ] remain to be processed, then, in step (1726), the breath- held-segment pointer k is incremented, and the noise detection process 1700 is repeated beginning with step (1704) for the next segment of breath-held sampled auscultatory sound data S m [io[k] : ii[k]]. Otherwise, from step (1724), the noise detection process 1700 terminates with step (1728).
  • the results from the docking system 27 may be transferred to a server computer system 56, for example, for subsequent transfer to an external data storage or processing system 58, for example, to provide for either viewing or analyzing the results at a remote location.
  • a server computer system 56 for example, for subsequent transfer to an external data storage or processing system 58, for example, to provide for either viewing or analyzing the results at a remote location.
  • the composite set of blocks of breath-held auscultatory-soundsensor time-series data S are screened prior to analysis by the associated Data Analysis Application (DAA) 54, for example, by first segmenting the set of blocks of breath-held auscultatory-sound-sensor time-series data S by heart cycle with an associated segmentation process 60, and then validating the associated heart cycle data with an associated heart cycle validation process 62, for example, to provide for additional screening for heart-phase associated noise.
  • DAA Data Analysis Application
  • an auscultatory sound signal preprocessing and screening process 2100 provides for preprocessing breath-held auscultatory sound signals 16.1 and an associated electrographic signal 37 from an ECG sensor 34, 34’, for example, as acquired, for example, in accordance with the above-described auscultatory-sound-sensing process 700, with concurrent recording of an associated electrographic signal 37 from an ECG sensor 34, 34’.
  • a good beat counter GB is initialized to zero, and a segment counter ksEG is initialized to a value of 1, wherein the good beat counter GB provides a count of the number of heart cycles for which the associated breath- held sampled auscultatory sound data S[] is not corrupted by either outliers or noise during diastole, and the segment counter ks EG is the value of the breath-held-segment pointer k for the current block B of breath-held sampled auscultatory sound data S[]
  • a segment of breath-held auscultatory sound signals 16.1 is acquired and preprocessed by an associated auscultatory sound signal acquisition and filtering process 2200, wherein, in step (2202), for each auscultatory sound sensor 12, in step (2204), a block of data of associated breath-held auscultatory sound signals 16.1 is acquired, for example, in accordance with the auscultatory-sound-sensing process 700, or as an associated segment k of a previously acquired and recorded breath-held sampled auscultatory sound data S m [io[k]: ii[k] for the m th auscultatory sound sensor 12.
  • the sampling rate may be reduced to 2 kHz.
  • step (2206) the breath-held auscultatory sound signal 16.1 is filtered by a fourth order Type II Chebyshev filter low-pass filter having a cut off frequency of 500 Hz to avoid aliasing, and then, in step (2208), decimated to a 2 kHz sampling rate using a phase invariant method that also involves low-pass filtering, so as to generate corresponding filtered-decimated breath-held sampled auscultatory sound signal 64.
  • the breath-held auscultatory sound signal 16.1 in some cases can include very low frequency - for example, around 10 Hz— vibrations that are believed to be associated with movements of the entire auscultatory sound sensor 12 stimulated by chest motion.
  • vibrations can be amplified by resonance characteristics of the tissue-sensor interface.
  • vibrations may decay very slowly, extending well into diastolic interval of heart beat, contaminating the signal of interest with relatively large amplitude unwanted interference. The net effect of such interference is an unstable signal baseline and distortion of the actual underlying heart sounds.
  • noise relevant to digitized acquisition of acoustic signals include: electric circuit thermal noise, quantization noise from the A/D converter, electro-magnetic interference, power line 60 Hz interference and acoustic noises relevant to human physiology and the environment where recording is done (ambient noise).
  • thermal noise power and A/D converter quantization noise are very low for the bandwidth of interest and may be significant only for the signals with amplitude in microvolt region.
  • recording artifacts may significantly reduce visibility of the signal of interest since these artifacts may have relatively high amplitude and may overlap in spectral content.
  • cardiac activity may also produce low frequency signals - for example, as a result of contraction of heart muscle, or as a result of valvular sounds— that may have valuable diagnostic information.
  • the spectrum of the artifacts may overlap with the acoustic spectrum of cardiac signals such as myocardium vibrations. Therefore, it can be beneficial to reduce baseline instability so as to provide for recording primarily acoustic signals originating from the cardiac cycle.
  • these very low frequency artifacts may be rejected by using additional signal filtering to suppress the associated characteristic vibration frequencies, for example, using a software-implement high-pass filter 66 having a 3 dB cut-off frequency above 10 Hz, for example, as provided for by a Savitzky-Golay-based high-pass filter 66’, wherein, in step (2210), the filtered-decimated breath-held sampled auscultatory sound signal 66 is smoothed by a Savitzky-Golay (SG) smoothing filter 68 to generate a smoothed breath-held sampled auscultatory sound signal 70, the laher of which, in step (2212), is subtracted from the filtered-decimated breath-held sampled auscultatory sound signal 64 to then generate the corresponding resulting high-pass-filtered breath-held sampled auscultatory sound signal 72 having a relatively-higher signal-to-noise ratio (SNR), but without significant distortion of the original filtered-decimated breath
  • SNR signal-to-noi
  • the digital Savitzky-Golay smoothing filter 68 is useful for stabilizing baseline wandering (for example, as may be exhibited in ECG signals) and provides for removing the low- frequency signal components without causing significant ringing artifacts.
  • the Savitzky-Golay smoothing filter 68 employs a least squares approximation of a windowed signal using a polynomial function.
  • the associated parameters of the Savitzky-Golay smoothing filter 68 include the window size M in samples that defines the associated cut-off frequency, and the polynomial degree N used for the approximation.
  • the associated roll-off range is fairly wide and the cut-off frequency is somewhat arbitrary.
  • the Savitzky-Golay smoothing filter 68 is defined by a least-squares fit of windowed original samples, with an N th degree polynomial,
  • the frequency response of the Savitzky-Golay smoothing filter 68 depends strongly on the window size M and polynomial order N.
  • the auscultatory sound signal acquisition and filtering process 2200 is repeated beginning with step (2202) for each of the auscultatory sound sensors 12, after which, from step (2216), the resulting blocks of high-pass-filtered breath-held sampled auscultatory sound signals 72— for each of the auscultatory sound sensors 12— are returned to step (2104).
  • a corresponding segment of a concurrent electrograhic signal 37 is acquired and preprocessed by an associated electrographic signal acquisition and filtering process 2200’, wherein, in step (2204’), a block of electrographic data 74 of the corresponding electrographic signal 37 from the ECG sensor 34, 34’ - in correspondence with the blocks of high-pass-filtered breath-held sampled auscultatory sound signals 72 returned by the above-described auscultatory sound signal acquisition and filtering process 2200 - is acquired, for example, in accordance with the above-described data acquisition process 800, or as an associated segment k of previously acquired and recorded electrographic data 74.
  • step (2206') the electrographic data 74 is filtered by a fourth order Type II Chebyshev filter low-pass filter having a cut-off frequency of 40 Hz, and then, in step (2208’), is decimated by a factor of ten, so as to generate corresponding filtered-decimated electrographic signal 76, which, in step (2216'), is returned to step (2104).
  • a fourth order Type II Chebyshev filter low-pass filter having a cut-off frequency of 40 Hz
  • the high-pass-filtered breath-held sampled auscultatory sound signals 72 or alternatively, the breath-held auscultatory sound signals 16.1 are segmented to identify the starting and ending points of each heart cycle, and to identify the starting and ending points of each associated diastole region or phase, thereof.
  • this may be accomplished using the breath-held auscultatory sound signals 16.1— or the corresponding associated high-pass-filtered breath- held sampled auscultatory sound signals 72 - alone, without relying upon the associated electrograhic signal 37 from the ECG sensor 34, 34’, or upon the corresponding associated filtered-decimated electrographic signal 76, for example, in accordance with the following portions of the disclosure and drawings of U.S. Patent No. 9,364,184: Abstract, FIGS. 1-50, Col. 1, Line 1 through Col. 3, line 59 (indicated), Col. 5, line 1 through Col. 34, line 55 (indicated), and the claims, which are incorporated herein by reference.
  • the electrographic signal 37 from the ECG sensor 34, 34’, and particularly, the corresponding associated filtered-decimated electrographic signal 76 responsive thereto provides an effective basis for segmenting the breath-held auscultatory sound signals 16.1, 72 by heart cycle, after which the high-pass-filtered breath- held sampled auscultatory sound signal 72 may then be used to locate the associated dominant SI and S2 heart sounds that provide for locating the associated heart phases of systole and diastole, data from the latter of which provides for detecting coronary artery disease responsive to information in the breath-held auscultatory sound signals 16.1, 72.
  • a heart-cycle segmentation and heart-phase identification process 2300 is called from step (2106) of the auscultatory sound signal preprocessing and screening process 2100 to locate the heart cycles responsive to the filtered-decimated electrographic signal 76, to then segment the high-pass- filtered breath-held sampled auscultatory sound signal 72 by heart cycle responsive thereto, and to then identify the region of diastole within each heart cycle (and implicitly, to therefor also identify the remaining region of systole) from an analysis of the high-pass-filtered breath-held sampled auscultatory sound signal 72 alone.
  • the normal human cardiac cycle consists of four major intervals associated with different phases of heart dynamics that generate associated audible sounds: 1) the first sound (SI) is produced by closing of mitral and tricuspid valves at the beginning of heart contraction, 2) during the following systolic interval, the heart contracts and pushes blood from ventricle to the rest of the body, 3) the second sound (S2) is produced by the closing of aortic and pulmonary valves and 4) during the following diastolic interval, the heart is relaxed and the ventricles are filled with oxygenated blood.
  • each QRS complex 78 illustrated in FIG. 25 is generated by an associated ventricular depolarization that precedes cardiac contraction.
  • the SI heart sound resulting from the closing of mitral and tricuspid valves is observed immediately after R-peak
  • the S2 heart sound resulting from the closing of aortic and pulmonary valves is observed at the end of T-wave of the ECG cycle.
  • This relative timing of the heart sounds SI, S2 and the QRS complex 78 of the associated electrographic signal 37, 76 provides for using the electrographic signal 37 to locate the beginning and end of each heart cycle without relying upon the breath-held auscultatory sound signal 16.1, 72 to do so.
  • the SI heart sound - marking the start of systole - follows shortly after the R-peak
  • the S2 heart sound - marking the end of systole and the beginning of diastole - follows thereafter, with diastole continuing until the next R-peak.
  • the electrographic signal 37, 76 may be distorted by low frequency baseline wandering, motion artefacts and power line interference.
  • a Savitzky-Golay-based high-pass filter 66’ - similar to that used in steps (2210/2212) when filtering the breath-held auscultatory sound signal 16.1 - may be used to cancel low-frequency drift, for example, prior to the subsequent low-pass filtering, in step (2206’), by the fourth order Type II Chebyshev filter low-pass filter, for example, having a 40 Hz cut-off frequency, and/or prior to decimation of the sampling by factor of 10 in step (2208’), which together provide for both reducing high frequency noise and emphasizing QRS complexes 78.
  • the heart- cycle segmentation and heart-phase identification process 2300 commences in step (2302) by calling an electrographic segmentation process 2600 that provides for identifying the locations of the R-peaks. More particularly, in step (2602), a block of electrographic data x[] is received, and, in step (2604), normalized so as to have a maximum absolute value of unity. Then, referring also to FIGS.
  • step (2702) of an associated first envelope generation process 2700 the block of filtered-normalized electrographic data x[] is used to generate an associated electrographic envelope waveform 80, Fs[] that emphasizes the R-peaks of the filtered-normalized electrographic data x[], wherein, for each value of index k — set in step (2702)— the corresponding value of the electrographic envelope waveform 80, Fs[k] is set in step (2704) responsive to a sum of values within a sliding window containing a subset of Nw points, as follows: Equation (9) is similar to a Shannon energy function, but with the associated discrete signal values raised to the fourth power - rather than a second power— to emphasize difference between R-peaks 80’ and baseline noise.
  • the value of the electrographic envelope waveform 80, Fs[] is calculated for each of the NPTS values of index k until, in step (2706), all points have been calculated, after which, in step (2708), the electrographic envelope waveform 80, Fs[] is returned to step (2606) of the electrographic segmentation process 2600.
  • the peaks of the electrographic envelope waveform 80, Fs[] are located, for example, by finding indices k for which the associated value of the electrographic envelope waveform 80, Fs[k] is a local maximum exceeding a fixed, predetermined threshold.
  • FIG. 28 illustrates the electrographic envelope waveform 80, Fs[] and associated R-peaks 80’ thereof, for a portion of a 10-second recording of a filtered-decimated electrographic signal 76.
  • the electrographic segmentation process 2600 for finding R-peaks is quite simple and stable when signal-to-noise ratio (SNR) of recording is sufficiently high. Otherwise, when the electrographic data 74 is excessively noisy, additional signal processing such as discrete wavelet decomposition may be used to further enhance R-peaks and facilitate QRS detection. Occasional relatively high amplitude transients due to patient movements or ambient interference may produce false peaks unrelated to QRS complex, and therefore complicate segmentation.
  • SNR signal-to-noise ratio
  • step (2610) the R-peaks 80’ of the electrographic envelope waveform 80, Fs[] are then validated by a peak validation process 2900, both with respect to temporal location - i.e. relative spacing— and magnitude - i.e. deviation from a median value,— so as to verify that the detected R-peaks 80’ correspond to R-peaks 80’ rather than noise.
  • step (2902) for each of the detected R-peaks 80’, in step (2904), if the difference in temporal locations of each pair of R-peaks 80’ is less than a time threshold TMIN, or if, in step (2906), the difference between the magnitude P(tpEAK(i)) of each R-peak 80’ and the median value of the magnitudes of all R-peaks 80’ is greater than a magnitude threshold VMAX, then, in step (2908), a status indicator associated with the R-peak 80’ being tested is set to IGNORE, in order to prevent the associated R-peak 80’ from being used for subsequent segmentation of the breath-held auscultatory sound signal 16.1, 72.
  • the peak validation process 2900 terminates in step (2912) and returns to step (2612) of the electrographic segmentation process 2600, which, in turn, returns the peak locations tpEAK[] to step (2302) of the heart-cycle segmentation and heart-phase identification process 2300, wherein valid R-peaks 80’ satisfy both of the following conditions:
  • tp EAK (i) is the time of the 1 th R-peak 80’ and R( ⁇ rEAk( ⁇ )) is the corresponding magnitude of the R-peak 80’.
  • the R-peaks 80’ of the electrographic envelope waveform 80, Fs[] are located in step (2608) by an associated R-peak detection process 5400 that provides for discriminating valid R-peaks 80’ from noise or other anomalies.
  • the R-peak detection process 5400 determines the peak values and associated samples times of the electrographic envelope waveform 80, Fs[] within a plurality of samples window 81, each of window width tw, and each offset from one another with an associated hop period t HOP .
  • the R-peak detection process 5400 commences with step (5402) by setting a window start time t MiN to point to the beginning of the electrographic envelope waveform 80, Fs[], and by initializing a window counter Nw to a value of 1. Then, in step (5404), the maximum value of the electrographic envelope waveform 80, Fs[] is determined within the sample window 81 extending between sample times tvu N and tvH N + tw, and stored in element Pw[Nw] of window-peak array Pw[], wherein, for example, the window width tw is sufficiently wide to span a single heart cycle, for example, with tw equal to about 1.5 seconds.
  • step (5406) the corresponding sample time at which the associated peak value occurred is stored as a corresponding element tp EAK _w[Nw] of an associated window- peak-time array tpEAK_w[] ⁇
  • step (5408) the window start time tMi N is incremented by the hop size t HOP , and, in step (5410), if the prospective end tvu N + tw of the prospective next sample window 81 is before the end t END of the electrographic envelope waveform 80, Fs[], then, in step (5412), if the most-recently detected R-peak 80’ is unique, then, in step (5414), the window counter Nw is incremented.
  • the R-peak detection process 5400 repeats beginning with step (5404).
  • the hop size t HOP is set to half the shortest heart cycle, but no less than 0.3 second, or a corresponding fraction of the window width tw, although it should be understood that the particular values of the window width tw and hop size t HOP are not limiting.
  • step (5412) provides for ensuring that each of the R-peaks 80’ in the window-peak array Pw[] are unique, and that the associated sample times in the window-peak array Pw[] are in monotonically increasing order.
  • a minimum peak- threshold PMINLIM is set equal to about 60 percent of the median amplitude of the R-peaks 80’ within the window-peak array Pw[]
  • a maximum peak-threshold PMAXLIM is set equal to twice the median amplitude of the R-peaks 80’ within the window-peak array Pw[]
  • step (5422) if the value of the second window counter iw is less than the number of windows Nw determined above in steps (5402) through (5414), then, in step (5424), if the magnitude of the currently -pointed-to R- peaks 80’, i.e.
  • Pw[iw] is greater than the minimum peak-threshold PMINLIM and less than the maximum peak-threshold PMAXLIM - indicating a potentially valid R-peak 80’ - then, in step (5226), the peak counter NPEAK is incremented, and the corresponding magnitude and time of the associated R-peak 80’, i.e. Pw[iw] and tpEAK_w[iw], are stored as respective values in in a corresponding peak array P [NPEAK] and peak-time array ⁇ REAK[NREAK] (alternatively, the window-peak array Pw[] and the window-peak-time array tpEAK_w[] could be reused in place for storing these values).
  • step (5424) provides for ignoring occurrences of noise and a noise or spurious signal—the magnitudes of each of which are less than the minimum peak-threshold PMINLIM, -- and provides for ignoring the example of a false R-peak 80’ that resulted from electro-static discharge (ESD), and which exceeded the maximum peak-threshold PMAXLIM .
  • step (5430) the second window counter iw is incremented to point to the next entry of the window-peak-time array tp EAK _w[], and the R-peak detection process 5400 repeats with step (5422). If, from step (5422), all entries of the have been processed, then, in step (5432), the process returns control to step (2608) of the electrographic segmentation process 2600.
  • a second aspect of a peak validation process 5600 provides for validating the R-peaks 80’ detected by the second aspect R-peak detection process 5400. More particularly, referring also to FIG.
  • step (5604) the associated heart-cycle period DT[IPEAK] between each pair of temporally-adjacent R-peaks 80’ is calculated in step (5604), indexed by a second peak counter ⁇ REAK that is initialized to a value of 1 in step (5602) and incremented in step (5608) so long as, in step (5606), the value thereof is less then NPEAK-1 that had been determined in steps (5424) and (5426) of the second aspect R-peak detection process 5400. Then, in step (5610), a minimum heart-cycle-period-threshold
  • DTMINLIM is set to 0.3 seconds (corresponding to a heart rate of 200 BPM).
  • a corresponding heart-rate HR[] array is calculated from the heart-cycle period DT[] array, wherein each element of the former is given by the inverse of the corresponding element of the latter, after which the median heart-rate HRMEDIAN is determined a the median of the values of the elements of the heart-rate array HR[]
  • a status STATUS[] array to initially indicate that each R-peak 80’ associated with the heart-rate array HR[] is“OK”, and the second peak counter IPEAK that is initialized to a value of 1.
  • step (5618) the heart-cycle period DT [IPEAK] is not greater than the minimum heart-cycle-period-threshold DTMINLIM, or, in following step (5620), the heart-rate HR[] is not both greater than 60 percent of the median heart-rate HRMEDIAN AND less than 140 percent of the median heart-rate HRMEDIAN, then, in step (5622) the STATUS [IPEAK] of the corresponding R-peak 80’ is set to “IGNORE”.
  • step (5624) if, in step (5624), all R-peaks 80’ have not been processed, then, in step (5626), the second peak counter IPEAK is incremented, and the second aspect peak validation process 5600 repeats with step (5618). Otherwise, from step (5628), the process returns control to step (2610) of the electrographic segmentation process 2600.
  • step (2304) for each auscultatory sound sensor 12, in step (2306), the associated high-pass-filtered breath-held sampled auscultatory sound signal 72 thereof is segmented based upon the locations of the associated R-peaks 80’ in the corresponding filtered-decimated electrographic signal 76.
  • FIG. 30 illustrates a correspondence between the breath-held auscultatory sound signal 16.1, 72 and the corresponding electrographic envelope waveform 80, Fs[], wherein the locations of the R-peaks 80’ in the latter are used to segment the former by heart cycle.
  • FIG. 30 illustrates a correspondence between the breath-held auscultatory sound signal 16.1, 72 and the corresponding electrographic envelope waveform 80, Fs[], wherein the locations of the R-peaks 80’ in the latter are used to segment the former by heart cycle.
  • FIG. 31 illustrates a stack of segments of the high-pass-filtered breath-held sampled auscultatory sound signal 72, each segment corresponding to a different corresponding heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9 - each resulting from a single corresponding heart beat— within the corresponding continuous high-pass-filtered breath-held sampled auscultatory sound signal 72 illustrated in FIG. 30.
  • first SI and second S2 heart sounds produced by the respective closing of atrioventricular and semilunar valves is somewhat ambiguous, and can be particularly difficult to locate in situations when the associated breath-held auscultatory sound signals 16.1 are recorded in an environment with relatively high level ambient noise, or if there are associated heart murmurs resulting from turbulent blood flow.
  • first SI and second S2 heart sounds of the cardiac cycle remain the most prominent acoustic features.
  • step (2306) following segmentation of the high-pass-filtered breath-held sampled auscultatory sound signal 72 into corresponding resulting associated heart-cycle segments in step (2306) — each comprising one heart cycle 82, resulting from a single corresponding heart beat, - beginning with step (2308), for each of the heart cycles 82 located in step (2306), in step (2310) a Shannon energy envelope is generated from the corresponding associated portion of the high-pass-filtered breath-held sampled auscultatory sound signal 72 in accordance with a corresponding second envelope generation process 3200, which is similar to the above-described first envelope generation process 2700 except for the associated envelope function. More particularly, referring to FIG.
  • step (3202) of the second envelope generation process 3200 for the k th of NSAMPLES samples of a block of high-pass- filtered breath-held sampled auscultatory sound data 72, s[] of the corresponding associated heart cycles 82, a corresponding value of an associated acoustic envelope waveform 84, Es[k] is generated in step (3204), responsive to a sum of values within a short-time sliding window - with different windows overlapping one another— containing a subset of Nw points, as follows in accordance with a canonical Shannon energy function Es[k] that provides a measure of the energy of the underlying high-pass-filtered breath-held sampled auscultatory sound signal 72, s[] :
  • the value of the acoustic envelope waveform 84, Es[k] is calculated for each of the NSAMPLES values of index k until, in step (3206), all points have been calculated, after which, in step (3208), the acoustic envelope waveform 84, Es[k] is returned to step (2310) of the heart-cycle segmentation and heart-phase identification process 2300.
  • the associated signal x[], s[], from which the corresponding envelope Fs[], Es[] is generated is zero-padded at the beginning and at the end with (Nw - l)/2 zeros, assuming Nw is an odd integer, wherein the first data point in the padded signal x[], s
  • FIG. 33 illustrates an example of an acoustic envelope waveform 84, Es[] in correspondence with a rectified version 86 (i.e.
  • the locations of the envelope peaks 88 S1 , 88 s2 associated with the corresponding SI and S2 heart sounds are identified using an adaptive threshold method that iteratively adjusts an associated threshold value down from a maximum value until the two most prominent envelope peaks 88 S1 , 88 s2 emerge within the associated heart cycle 82, above the a particular threshold limit.
  • the final position of envelope peaks 88 S1 , 88 s2 associated with the corresponding SI and S2 heart sounds is found as the apex of a local quadratic model 90.1, 90.2 of the acoustic envelope waveform 84, Es[], as described more fully hereinbelow.
  • Adaptive thresholding provides for accommodating substantial variation in the relative magnitude of the envelope peaks 88 S1 , 88 s2 that might occur either from one heart beat to another, from one patient to another, or from one recording site to another.
  • step (2316) the locations of the envelope peaks 88 S1 , 88 s2 associated with the corresponding SI and S2 heart sounds are validated using a normalized acoustic envelope waveform 84, Es[], i.e. normalized to a range of 0 and 1, and the associated local quadratic models 90.1, 90.2 thereof, in accordance with a minimum-time-spacing criteria used to remove or ignore spurious transient peaks unrelated to heart sounds, similar to equation 10 above that is associated with step (2904) of the above-described peak validation process 2900 used to validate the electrographic envelope waveform 80, Fs[].
  • the acoustic envelope waveform 84, Es[] is searched relative to the associated indices ksi_ PEAK , ks 2 _pn ⁇ i ⁇ — respectively associated with the corresponding respective envelope peaks 88 S1 , 88 s2 — to find adjacent data points therein,— i.e. having associated indices ksi-, ksi + , ks2-, ks2 + — for which the corresponding values of the acoustic envelope waveform 84, Es(ksi-), Es(ksi+), Es(ks2-), Es(ks2+) are each about 5 percent down, i.e. 95 percent of, the corresponding values Es(ksi PEAK), Es(ks2_PEAK> of the associated envelope peaks 88 S1 , 88 s2 .
  • step (2320) respective local quadratic models ESl(k), 90.1’ and ES2(k), 90.2’ are fitted - for example, by least-squares approximation— to the three points associated with each of the corresponding respective envelope peaks 88 S1 , 88 s2 as follows:
  • ES l (k) Quadratic Fit ( ⁇ ksi-, Es(ksi-) ⁇ , ⁇ ksi_PEAK, Es(ksi_PEAK) ⁇ , ⁇ ksi+, Es(ksi+) ⁇ ) (13a)
  • ES2(k) Quadratic Fit ( (ks2-, Es(ks2-) ⁇ , ⁇ ks2_PEAK, Es(ks2_PEAK) ⁇ , ⁇ ks2+, Es(ks2+) ⁇ ) (13b) Then, referring again to FIG.
  • S3. S4 or S5 sounds may be present in the heart cycles 82 for some test-subjects 22.
  • the purpose of identifying SI, S2, S3, S4. and S5 regions is to help extract or derive useful features (such as Indicators that correlate highly with the existence of coronary artery disease (CAD)) from heart sounds.
  • useful features such as Indicators that correlate highly with the existence of coronary artery disease (CAD)
  • CAD coronary artery disease
  • these features can be used in direct detection and machine learning for discriminating CAD patients from non-CAD patients.
  • the prerequisite of many feature extractions is to correctly identify the SI and S2 sounds.
  • the S3, S4, and S5 sounds do not appear for every patient. But each may indicate one or more cardiac conditions, such as, CAD, Aortic Insufficiency, Aortic Stenosis, Luminal Irregularities, and Mitral Regurgitation.
  • the regions within diastole during which the S3 and S5 sounds may occur are located relative to the diastasis region of diastole, which is a relative rest period of the heart during mid-to-late diastole, during which period the heart motion is minimal.
  • the region during which the S4 sound may occur is located relative to the R-peak 80’ at the end of the heart cycle 82 and the beginning of the next heart cycle 82.
  • the starting point of the diastasis region is determined using what is referred to as the Weissler or Stuber formula for the period of delay DT— in milliseconds— from an R-peak 80’ to the starting point of the diastasis region, given by the following:
  • ATR-R is the time interval in milliseconds between R-peaks 80’.
  • this is approximately the ending point for the region most likely to include S3, i.e. the S3 region.
  • the starting point for the S3 region is determined by advancing relative to the starting point of diastasis - or, equivalently, the end point of the S3 region— by a time interval ATs3.
  • the time interval commences at about 100 milliseconds prior to the starting point of diastasis and ends at the starting point of diastasis.
  • the time interval of the S3 region is taken to extend from about 60 milliseconds prior, to 60 milliseconds after, the starting point of diastasis.
  • the S3 swing is calculated by subdividing the S3 region of the associated breath-held auscultatory sound signal 16.1, for example, band-pass-filtered breath-held sampled auscultatory sound data s[] having a pass-band in the range of 20, 25, or 30 Hz to 40 or 50 Hz, , into a series of - i. e. one or more— time intervals, and calculating or determining one or more of the difference between the maximum and minimum amplitude values— i.e. maximum amplitude - minimum amplitude,— the minimum amplitude, and the maximum amplitude, for each interval in the S3 region.
  • the associated breath-held auscultatory sound signal 16.1 is also analyzed in the frequency domain using a Short-Time Fourier Transform (STFT), for example, in one set of embodiments, having a 1 Hz frequency resolution and a 0.0025 second time resolution— but generally, using frequency and time resolutions that may be empirically adjusted to improve detection and discrimination - in cooperation with an associate windowing method, for example, using Chebyshev windowing on a sliding window that is moved along the S3 region of the breath-held auscultatory sound signal 16.1.
  • STFT Short-Time Fourier Transform
  • the S3 swing values and frequency-domain features are saved for further calculations and/or for use as an input to one or more of the below-described classification processes.
  • an unsupervised clustering method is applied on all the generated features to classify heart beats into two clusters that respectively include“with S3” and“without S3 heart beats.
  • S3 is analyzed on a beat-by-beat basis. Given the S3 is not an intermittent signal, one strategy is to analyze all heart beats from a given patient, and if S3 appears in more than one third of all the heart beats, that patient would be identified as having S3
  • There are some patients who exhibit a low-ejection fraction ratio e.g. an ejection fraction (E.
  • the S4 region is located in a final portion of diastole, starting at the P wave on the ECG signal.
  • an approximate starting point for the S4 region may be used, for example, a time interval of ATs4, for example, from 10 to 20 percent of the period of the heart cycle 82, for example, about 100 to 200 milliseconds in a 1 second duration heart cycle 82, in advance of the R-peak 80’ at the end of the heart cycle 82, or equivalently, in advance of the beginning of the SI sound of the next heart cycle 82.
  • the band-pass-filtered breath-held sampled auscultatory sound data s[] having a pass-band in the range of 20, 25, or 30 Hz to 40 or 50 Hz is subdivided into a series of intervals and the associated S4 swing within each interval is calculated as the absolute value of the difference between maximum and minimum amplitude magnitudes of the raw or high-pass-filtered data within that interval in the S4 region.
  • the S4 swing is calculated separately for audible (above 20 Hz) and inaudible (below 20 Hz) frequency ranges, for which the configurations of the associated signal filters are specific to the particular frequency range.
  • the minimum and maximum values of the signal will depend upon the associated frequency range of the associated signal, and the type of filter used to generate the filtered signal.
  • the S2 swing is also similarly calculated over the associated S2 region, in addition to calculating the S4 swing as described hereinabove.
  • the ratio of the S4 swing to S2 swing provides a measure of the likelihood that a patient exhibits an S4 sound, with that likelihood increasing with increasing value of the S4-to-S2 swing ratio.
  • FIG. 71 illustrates a heartbeat for which the S4-to-S2 swing ratio has a value of 0.49 for the illustrated channel 1 of the band-pass- filtered breath-held sampled auscultatory sound data s[].
  • the S2 and S4 swing values, and/or the S4-to-S2 swing ratio are saved for further calculations and/or for use as an input to one or more of the below-described classification processes.
  • the mean value of the S4-to-S2 swing ratio is calculated for an entire population of patients, beat by beat, and then the median of S4-to-S2 swing ratio is taken across all beats of each patient.
  • the S5 region is identified as the end of the S3 region to the start of the S4 region. Accordingly, the starting point is determined using the above-described Weissler or Stuber formula. As mentioned, this is approximately the ending point for the S3 region.
  • the ending point of the S5 region is located at the beginning of a time interval of A TSJ. for example, about 100 milliseconds, in advance of the R-peak 80’ at the end of the heart cycle 82, or equivalently, in advance of the beginning of the SI sound of the next heart cycle 82.
  • the S5 swing is calculated by subdividing the S5 region of the associated breath-held auscultatory sound signal 16.1, e.g.
  • the S5 swing values may be saved for further calculations and/or for use as an input to one or more of the below-described classification processes.
  • step (2324) If, in step (2324), all heart cycles 82 in the high-pass-filtered breath-held sampled auscultatory sound signal 72 have not been processed, then the heart-cycle segmentation and heart-phase identification process 2300 is repeated beginning with step (2308) for the next heart cycle 82. Otherwise, from step (2326), if all auscultatory sound sensors 12 have not been processed, then the heart-cycle segmentation and heart-phase identification process 2300 is repeated beginning with step (2304) for the next auscultatory sound sensor 12.
  • the heart-cycle segmentation and heart-phase identification process 2300 returns either the mean values ksi, ks2 of the corresponding root locations ⁇ ksi sTART, ksi END ⁇ , ⁇ ks2_START, ks2_END ⁇ associated with the corresponding SI and S2 heart sounds, or the corresponding mean values tsi, ts2 of the associated times, i.e.
  • segmented high-pass-filtered breath-held sampled auscultatory sound data 72, s[] from each of the heart cycles 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9 may be synchronized with respect to the mean time ts2 associated with the S2 heart sound of each heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9, as illustrated in FIG.
  • 35 may be synchronized with respect to the mean time tsi associated with the SI heart sound of each heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9, wherein the region of diastole extends from the time associated with the second root ks2_ END , i.e. t(ks2_ END ), associated with the S2 heart sound, to the end of the associated heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9.
  • either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72 is first analyzed within each region of diastole for each heart cycle 82 identified in step (2106/2300), and for auscultatory sound sensor 12, m to identify any outliers, and to detect, for selected pairs of auscultatory sound sensor 12, excessive noise.
  • a heart-cycle pointer ku EAT is first initialized to a value of 1, after which, via step (2108), for each heart cycle 82 and for each auscultatory sound sensor 12, m, in step (2110), each region of diastole of either the breath-held auscultatory sound signal 16.1, or the corresponding high- pass-filtered breath-held sampled auscultatory sound signal 72, i.e. extending between samples is2[ksEG, / ( BEAT] and isifksEG, keEAT+l] is checked for outliers, e.g.
  • heart-cycle pointer knFA r points to the particular heart cycle 82 within the selected breath-held segment ks EG , and is 2 [ks EG , kii EA r] corresponds to the ks2_ END of the associated heart cycle 82 at the beginning of diastole, and isifks EG , k BEAT+i ] corresponds to the end of diastole of that heart cycle 82, and the beginning of the systole for the next heart cycle 82.
  • the heart-cycle pointer knFA r points to the particular heart cycle 82 within the selected breath-held segment ks EG , and is 2 [ks EG , kii EA r] corresponds to the ks2_ END of the associated heart cycle 82 at the beginning of diastole, and isifks EG , k BEAT+i ] corresponds to the end of diastole of that heart cycle 82, and the beginning of
  • the beat outlier detection process 3600 commences in step (3602) with receipt of the identifier m of the auscultatory sound sensor 12, the associated sampled auscultatory sound data S m [], the segment counter ks EG , and the heart-cycle pointer 13 ⁇ 4EAT. Then, in steps (3604) and (3606), a standard-deviation counter ksTDDEv is initialized to a value of 1, and an index JMIN is set equal to is2
  • the standard deviation array STD[] is rank ordered, and the median value thereof, MedianSTD, is used in step (3620) to calculate a standard deviation compactness metric STDDEVCM, as follows:
  • step (3622) if the standard deviation compactness metric STDDEVCM exceeds a threshold, for example, in one set of embodiments, equal to 6, but generally between 1 and 10, the particular region of diastole for the particular breath-held segment from the particular auscultatory sound sensor 12, m, is flagged as an outlier in step (3624). Then, or otherwise from step (3622), in step (3626), the process returns to step (2110) of the auscultatory sound signal preprocessing and screening process 2100.
  • a threshold for example, in one set of embodiments, equal to 6, but generally between 1 and 10.
  • step (2112) of the auscultatory sound signal screening process 2100 if an outlier was detected in step (3622) and flagged in step (3624) of the beat outlier detection process 3600, then in step (2120), if the end of the breath-held segment has not been reached, i.e. if 13 ⁇ 4EAT ⁇ NBEATs(ksE G ), then the auscultatory sound signal preprocessing and screening process 2100 repeats beginning with step (2108) for the next heart cycle 82.
  • the above-described noise detection process 1700 may be called from step (2114) to provide for determining whether or not a particular region of diastole for a particular heart cycle 82 of either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72, is corrupted by excessive noise, and if so, provides for flagging that region of diastole as being excessively noisy so as to be prospectively excluded from subsequent processing, wherein above references to breath-held auscultatory sound signal 16.1 in the description of the associated noise detection process 1700 should be interpreted as referring to the corresponding region of diastole, i.e.
  • step (2120) if a noise threshold was exceeded in step (1930) of the noise-content-evaluation process 1900, in step (2120), if the end of the breath-held segment has not been reached, i.e. if ke EAT ⁇ N BEAT s(ks EG ), then the process repeats beginning with step (2108) for the next heart cycle. Otherwise, from step (2116), the good beat counter GB is incremented in step (2118) before continuing with step (2120) and proceeding therefrom as described hereinabove.
  • step (2122) if the end of the breath-held segment has been reached, in step (2122), if a threshold number of valid heart cycles has not been recorded, the process repeats with step (2104) after incrementing the segment counter ks EG in step (2123). Otherwise, the recording process ends with step (2124).
  • each breath-holding interval B, ks EG of either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72 is segmented into individual heart beats 82 (i.e. heart cycles 82, wherein reference to heart beats 82 is also intended as a short-hand reference to the associated breath-held sampled auscultatory sound data S[]) and the diastolic interval D is analyzed to determine the associated noise level to provide for quality control of the associated breath-held sampled auscultatory sound data S[], so as to provide for signal components thereof associated with coronary artery disease (CAD) to be detectable therefrom.
  • CAD coronary artery disease
  • Quality control of the recorded signals provides for detecting weak signals that may indicate health problems but can otherwise be blocked by strong noise or unwanted interference.
  • the present method is developed for quantitative control of signal quality and can be deployed in the recording module 13 for real time quality monitoring or can be used at the post recording stage to extract only low-noise heart beats 82 that satisfy specific quality condition.
  • a second aspect of an auscultatory sound signal preprocessing and screening process 3700 also referred to as a beat selection algorithm, as was the first aspect auscultatory sound signal preprocessing and screening process 2100 - provides for collecting heart beats 82 that have diastolic noise level below the specific mean noise power level threshold Po.
  • the associated auscultatory-sound-sensing process 700 proceeds through sequential acquisition of heart sound intervals with duration between 5 and 15 sec (for example, breath-holding intervals B).
  • step (3702) the sampled auscultatory sound data S m [], the identifier m of the associated auscultatory sound sensor 12, and the associated segment counter ksEG are received, and, in step 3704), the corresponding recorded data block is passed through the heart-cycle segmentation and heart-phase identification process 2300 that identifies the beginning and the end of each heart beat 82 using synchronized ECG recording or another signal processing code that identifies timing of the first (SI) and second (S2) heart sounds, as described hereinabove.
  • the second aspect auscultatory sound signal preprocessing and screening process 3700 input parameters also include the mean noise power level threshold Po and the required number NG MIN of high quality heart beats 82.
  • a good beat counter GB is initialized to a value of 0.
  • a two-dimensional array of heart beats 82 is created, for example, as illustrated in FIGS. 31 and 35.
  • the heart-cycle segmentation and heart-phase identification process 2300 also identifies timing of diastolic interval D of each heart beat 82.
  • the breath- held sampled auscultatory sound data S[] is normalized with respect to absolute value of the S2 peak, and, in accordance with a first aspect, in step (3712), the mean noise power P is computed within a time window T w .
  • the time-window T w is slid along the diastole D with the 50% overlap to compute an array of localized signal power P[]
  • the maximum power PMAX of diastole D is determined i.e. as given by the maximum of the array of localized signal power P[], and, in step (3714), is compared against the mean noise power level threshold Po, the latter of which in one set of embodiments, for example, has a value of -20 dB.
  • a noise power level P[iw] in one of the time-windows T w greater than the mean noise power level threshold Po may indicate either excessive noise or presence of a large amplitude transient outlier.
  • the quality test in accordance with a second aspect of heart-beat quality assessment procedure, the quality test consists of two stages designed to detect high amplitude transient spikes and noisy beats with broad-band spectrum.
  • the quality tests are performed on the diastolic portion of the heart cycle 82, which is the region of primary interest for CAD detection.
  • the stride period is equal to the duration of the time-window T w , so that adjacent time-windows T w abut one another without overlap.
  • An outlier power threshold PLIM is determined by adding 6 dB to the median value of Pi for all time-windows T w , and in step (3714), if the value of Pi exceeds PLIM for any time-window T w , then, in step (3720), the current heart cycle 82 is ignored.
  • the associated noise power threshold Rt ⁇ is defined with respect to the 2-byte A/D converter range, so that:
  • Th 1 P ADC -1 1 0 u (18)
  • Th is a predetermined threshold, for example, -50 dB
  • PADC is the power associated with the maiximum signed range of a 2-byte A/D converter, i.e. (32767) 2 . Accordingly, if, in step (3714), the mean diastole power Pm exceeds the noise power threshold Pn, then, in step (3720), the current heart cycle 82 is ignored.
  • step (3720) If, from step (3714), the diastolic signal power exceeds the mean noise power level threshold Po, then, in step (3720), the associated heart beat 82 is labeled as noisy beat and is not counted in the overall tally of heart beats 82. Otherwise, from step (3714), the good beat counter GB in step (3716), and if, in step (3718), if the number of good heart beats 82, i.e. the value of the good beat counter GB, is less than the required number NGMIN of high quality heart beats 82, then the second aspect auscultatory sound signal preprocessing and screening process 3700 repeats with step (3708) for the next heart cycle 82. In one set of embodiments, if the required number of high quality heart beats 82 is not reached within a reasonable period of time, then the user is informed that recording is excessively noisy so that additional actions can be performed to improve signal quality.
  • the sampled auscultatory sound data S m [] is further preprocessed to emphasize acoustic signals in diastole D and extract signal specific features that can be used for disease classification.
  • the mean position of the peak of the S2 heart sound - for example, the above-described mean value ts2 -- is determined for each heart beats 82, after which the heart beats 82 are synchronized with respect thereto, for example, as illustrated in FIGS.
  • 39a and 39b which respectively illustrate a stack of heart beats 82 before and after this S2 alignment, wherein the synchronization of the heart beats 82 with respect to the S2 heart sound helps to locate an acoustic signature that might be present in a majority of the heart beats 82 and that is coherent between or amongst different heart beats 82.
  • the heart beats 82 are normalized with respect to time so as to compensate for a variation in heart-beat rate amongst the heart beats 82, and to compensate for a resulting associated variation in the temporal length of the associated diastolic intervals.
  • heart-beat rate is always changing and typically never remains the same over a recording period of several minutes, which if not compensated, can interfere with the identification of specific signal features in diastole D, for example, when using the below-described cross-correlation method.
  • heart-beat segmentation alone provides for aligning heart-beat starting points, variations in the heart-beat rate can cause remaining features of the heart cycle 82 to become shifted and out of sync - i.e.
  • offsets can be removed if the associated heat beats 82 are first transformed to common normalized time scale t/T*, where T* is the fixed time interval, for example, the duration of the slowest heart beat 82, followed by beat resampling and interpolation so as to provide for normalizing the original signal at a new sampling rate, for example, as illustrated in FIGS. 40a and 40b that respectively illustrate stacks of heart beats 82 before and after such a time normalization, wherein FIG. 40a illustrates complete heart cycles 82, and FIG. 40b illustrates only the temporally -normalized diastolic portions D thereof.
  • a final stage of signal pre-processing provides for extracting acoustic features from the recorded heart beats 82.
  • feature extraction typically involves certain transformation of raw data to low-dimensional or sparse representation that uniquely characterizes the given recording.
  • the raw data can be transformed in a set of images and some image recognition algorithm like convolutional neural net can be employed for automatic feature selection.
  • image recognition algorithm like convolutional neural net can be employed for automatic feature selection.
  • a local cross- correlation (CC) of multiple heart beats 82 provides for identifying a relatively high-pitch signal component in the diastolic phase of the heart cycle 82 occurring in multiple heart beats 82, which can be result from turbulent blood flow, wherein pairs of heart beats 82 of a 2-D stack of heart beats 82 - each segmented from the above-described high-pass-filtered breath-held sampled auscultatory sound data 72, s[]— are cross-correlated with one another by computing an associated set of cross-correlation functions R xixj for each pair of heart beats 82, xi[], x j [].
  • This computation is made using a sliding short-time window with N w samples (for example, typically 128) which is advanced in time one sample per each iteration of the cross-correlation computation.
  • N w samples for example, typically 1228 which is advanced in time one sample per each iteration of the cross-correlation computation.
  • the cross-correlation is computed without time lag, resulting in an array - of the same size as that of the original signals— that is given by:
  • the cross-correlation assigned to each beat is given by an average of cross-correslations thereof with the remaining Nb - 1 heat beats 82:
  • xi and xj are the diastolic high-pass-filtered breath-held sampled auscultatory sound data 72, s[] of two distinct heart beats 82, and N b is the total number of heart beats 82 in the 2- D stack.
  • N b is the total number of heart beats 82 in the 2- D stack.
  • cross-correlation peaks associated with a micro-bruit signal occurring at approximately at the same temporal location within the diastole interval from one heart beat 82 to another will produce distinct bands across the image within the same temporal region of each heart beat 82, for example, as illustrated in FIG. 42, which illustrates an image of the cross-correlation function R xi as a function of n (corresponding to time) for each heart beat 82, with different heart beats 82 at different ordinate positions, wherein the value of the cross-correlation function is indicated by the color of the associated pixels.
  • the cross-correlation operation provides for emphasizing signal features that are coherent within the current time window, and provides for suppressing uncorrelated noise, thereby providing for increasing the associated signal-to-noise ratio.
  • acoustic features in diastolic interval of heart beat 82 can be visualized using continuous wavelet transform (CWT).
  • the wavelet transform processing is similar to the short-time cross-correlation but instead of cross-correlating signals from different heart beats 82, the signal of interest is correlated using a wavelet function with limited support to facilitate temporal selectivity.
  • Output of the wavelet transform is a two-dimensional time-frequency representation of signal power
  • An example of a wavelet transform - using a 6th order Morlet wavelet— is illustrated in FIG. 43, where the associated corresponding original color map represents distribution of the signal power in a time-frequency plane, with normalized time-shift - as a function of b— as the abscissa, and frequency as the ordinate, with frequency f in Hertz given as follows:
  • the wavelet representation can be reduced further by dividing the time axis into discrete intervals and computing overall signal power within such interval and specific frequency bandwidth.
  • the wavelet image may be subdivided in time intervals of 200 milliseconds and two frequency bands of 10 Hz - 40 Hz and 40 Hz - 100 Hz.
  • the resulting output vector of the signal power within the associated intervals of time and frequency can be used as an input to a neural network classifier.
  • step (3730) of the second aspect auscultatory sound signal preprocessing and screening process 3700 - which may also be used in cooperation with the above-described first aspect auscultatory sound signal preprocessing and screening process 2100 - incorporates one or more feature extraction algorithms which identify significant signal parameters that can be linked to coronary artery disease (CAD) and which can be used for training a machine learning algorithm for automatic CAD detection.
  • CAD coronary artery disease
  • each heart beat 82 might contain over 4000 samples per each of six channels.
  • Such large amount of highly correlated variables makes the usage of the raw waveform for classification very difficult without additional signal processing to reduce the dimensionality of the problem.
  • Such dimensionality reduction can be achieved by use of an appropriate feature extraction algorithm that identifies a reduced set of parameters that are related to CAD.
  • the feature extraction procedure provides a mapping of the high-dimensional raw data into the low-dimensional feature space with adequate inter-class separation.
  • standard dimensionality reduction routines such as singular value decomposition (SVD) or principal component analysis (PCA) may be used to decompose raw data onto orthonormal basis and to provide for selecting relevant features with minimal loss of information.
  • the time domain signal itself can be transformed prior to feature extraction to emphasize unique features thereof. For example, frequency domain representation by Fourier transform can be advantageous for feature extraction if the signal contains discrete set of characteristic frequencies.
  • the performance of a signal classifier can be dramatically improved by excluding a large number of irrelevant features from analysis.
  • the signal classification problem begins with a mapping from the original high-dimensional space (size N) to a feature space (size p « N), followed by a mapping of the feature space to an m-dimensional space, wherein the dimension m is equal to the number of classes.
  • a binary classification problem e.g. CAD or no CAD
  • m 2.
  • step (3730) of the second aspect auscultatory sound signal preprocessing and screening process 3700 employs a wavelet packet transformation (WPT) for sparse representation of heart sounds in time-frequency domain, followed by a custom designed binary classifier.
  • WPT wavelet packet transformation
  • Several standard classifier algorithms can be trained using reduced feature set, and to provide for binary classification of the associated heart sounds- useable either individually or in combination,— including , but not limited to, a support vector machine (SVM), a fully -connected artificial neural network (ANN), or a convolution neural network (CNN) applied to two-dimensional time-frequency images.
  • SVM support vector machine
  • ANN fully -connected artificial neural network
  • CNN convolution neural network
  • a wavelet packet transformation (WPT) processing stage provides for reducing dimensionality by converting raw time-domain auscultatory sound signals 16 into a time-frequency basis using discrete wavelet transform (DWT), followed by elimination of associated components that do not provide significant contribution to the original signal or do not provide substantial contrast between two classes of interest.
  • WPT wavelet packet transformation
  • DWT discrete wavelet transform
  • an initial (input) signal x[n] is passed through a series of stages at which low-pass and high-pass filter functions (quadrature mirror filters) are applied to obtain approximation ii j (k) and detail d j (k) coefficients at the /* level of decomposition.
  • the discrete wavelet transform is implemented using a Daubechies wavelet family, for example, a Daubechies 4 (db4) wavelet family, for which the associated scaling function f— having low- pass filter coefficients ho through h — is illustrated in FIG. 59, and for which the associated wavelet function Y— having high-pass filter coefficients go through - is illustrated in FIG. 60.
  • a Daubechies wavelet family for example, a Daubechies 4 (db4) wavelet family, for which the associated scaling function f— having low- pass filter coefficients ho through h — is illustrated in FIG. 59, and for which the associated wavelet function Y— having high-pass filter coefficients go through - is illustrated in FIG. 60.
  • an input time series W j.k fl/ from decomposition level j is transformed into two output time series at decomposition level /+/. i.e. W j+i,2k [l] and W j+i,2k+i /l/. each containing half the number of samples as in the input time series, and mutually-exclusive halves of the frequency content of the input time series, wherein the transformation to generate the Wj+i ⁇ k+ifl] lower-half bin of frequency content is given by the transformation 61.1g 61.2g illustrated in FIG.
  • the filter functions are designed to provide for energy conservation and lossless reconstruction of the original signal from the set of transformed time series W j.k fl/ from a particular decomposition level j. These properties along with smoothness requirements define the family of scaling and wavelet functions used for decomposition.
  • the wavelet packet transformation is generalization of the standard multi-level DWT decomposition, wherein both approximation and detail coefficients are decomposed using quadrature mirror filters, for example, as described in M. Wickerhauser,“Lectures on Wavelet Packet Algorithms”, http://citeseerx.ist.psu.edu, which is incorporated herein by reference.
  • FIG. 63 illustrates the bandpass frequency responses associated with each of the outputs thereof.
  • the wavelet packet transformation (WPT) decomposition can be described by the following recursive expressions:
  • WPT wavelet packet transformation
  • FIG. 65 illustrates a wavelet packet transformation (WPT) energy map of the high-pass-filtered breath-held sampled auscultatory sound signal 72 for the heart cycle 82 illustrated in FIG.
  • the total energy from all frequency bins is the same, i.e.
  • the wavelet packet transformation (WPT) energy map of FIG. 65 illustrates that most of the signal energy is concentrated in the relatively low frequency bands, and that only the levels 7 and 8 of decomposition show the fine structure of the energy distribution. Accordingly, this example shows that most of the wavelet packet transformation (WPT) nodes 92 can be ignored without significant loss of information.
  • the Wavelet Packet Transform provides a redundant representation of actual signal using the discrete wavelet transform with various decomposition levels that provide higher frequency resolution (in terms of filter banks) at the expense of reduced time resolution.
  • the WPT of the signal can be described by a set of coefficient wj(k, 1), where j is the decomposition level, k is the filter bank index and 1 is the coefficient index within filter bank.
  • An example of WPT decomposition is shown in FIG. 65 where each cell outlines positions of filter bank for each level and intensity is related to the signal energy from equation 24. It can be seen that energy distribution wi thing specific frequency bands becomes apparent at high decomposition levels ( j > 6). Although, the color of the high frequency cells is dark blue, the values of the associated WPT coefficients is not zero.
  • the entire WPT energy map is highly redundant and not well suited in its raw form for signal classification. Instead, only those cell that uniquely characterize a given signal are kept while ignoring remaining cells.
  • An example of best basis selection applied to the WPT map in FIG. 65 is shown in FIG. 66. Values the best basis would generally not coincide with the energy map shown in FIG. 65 because entropy and energy differ from one another,
  • FIG. 65 the total energy of a particular cell is represented by the intensity thereof.
  • FIG. 65 illustrates that most of the energy is concentrated in the relatively -lower frequency bands.
  • the cells of FIG. 66 - for which intensity is inversely related to entropy - are considered to be a signature of a given signal, and do not overlap with the corresponding cells of FIG. 65, the latter of which is representative of energy. None of the cells below the highlighted (children) cells in FIG. 66 are selected because they exhibit high entropy.
  • the wavelet packet transformation (WPT) energy map and the associated best basis selection can be used to reduce dimensionality of the heart beat classification problem by analyzing the signal represented by transformed time series W/ A /// and rejecting information irrelevant for the classification task.
  • the wavelet packet transformation (WPT) is one of a variety of signal processing techniques that can be used for extraction of important parameters or features suitable for prediction of CAD.
  • the very basic set of features may include typical metrics of raw signals (amplitude, timing, spectral power and other) that can be derived from segmented heart beats. However, such hand-crafted feature set may not be optimal for current problem of CAD classification.
  • the output of this data processing stage is a vector of p elements with p « N, where N is the size of raw signals.
  • the feature vector can be represented either as a 1-D array of p elements or as a 2-D matrix for a classification algorithm operating on image information.
  • classification algorithms include support vector machine (SVM), feed-forward artificial neural network (ANN) and convolutional neural network (CNN), which is particularly suitable for 2-D image classification.
  • a support vector machine is a powerful supervised machine leaning algorithm frequently used for data classification problems, which has several useful properties and can operate reliably on a small data sets with poor class separation.
  • the SVM classifier will create decision hyperplane in feature space with highest separation between two classes using training data set.
  • FIG. 67 illustrates a trained SVM classifier and decision boundary for a two-dimensional feature space with two classes. The support vectors are circled, and the highest margin hyperplane is shown by the solid straight line.
  • the tolerance of SVM algorithm to data misclassification can be tuned by adjusting an associated C parameter during training.
  • a nonlinear decision boundary can be created using a kernel function to project data into higher dimensions and apply SVM algorithm to modified data.
  • the SVM algorithm can be used as a CAD classifier with data recorded by the Data Recording Application (DRA) 14, 14.1.
  • DRA Data Recording Application
  • recordings Prior to sending data to SVM algorithm, recordings are processed by beat segmentation and feature extraction stages, to produce a feature vector for each test-subject 22, by preprocessing of n available recordings and extracting p features, with each channel data then transformed into an n x p feature matrix.
  • the feature vector extracted from the segmented beats can be either a set of custom selected metrics (amplitude, timing of specific segment, energy, statistic parameters, sample entropy and others) or a subset of wavelet packet transformation (WPT) coefficients associated with the signal region of interest.
  • WPT wavelet packet transformation
  • a principal component analysis (PCA) procedure can be applied to identify a subset of features with highest variance and eliminate correlating features.
  • PCA principal component analysis
  • the feature matrix is split into testing and training sets with ratio 1 to 4 respectively.
  • the training set is used to train SVM and optimize classifier hyper-parameters, while the testing set is used to evaluate classifier performance with unseen data.
  • Computer code that provides for implementing a SVM classifier is available in several open source packages for Python and R programming languages, for example, the skleam machine learning package in Python.
  • a feed-forward artificial neural network provides an alternative option for classification of auscultatory sound signals 16 following preprocessing and feature extraction stages.
  • An artificial neuron is a nonlinear processing element with p inputs and a nonlinear activation function that generates an activation output. Assuming x(p) is a feature vector which is sent to the p inputs of each neuron, then the associated activation output of i lh neuron is computed as:
  • the specific form of the activation function g is chosen at the design stage, wherein commonly used functions include sigmoid, hyperbolic tan and rectified linear unit (ReLu).
  • the network of interconnected neurons constitutes the artificial neural network (ANN), which can be capable of modeling relatively complicated relationships between the input vector and target class variables by adjusting weights and biases during training stage.
  • ANN artificial neural network
  • ANN artificial neural network
  • properties of a specific artificial neural network (ANN) implementation are defined at the design stage and include: 1) number of hidden layers, 2) number of neurons per hidden layer, 2) type of activation function, 3) learning rate and 4) regularization method to prevent overfitting.
  • the artificial neural network (ANN) is implemented using the open source TensorFlow deep learning framework, which provides for setting each of these parameters.
  • the neuron connection strength is defined by the weight matrix wy for each layer which is adjusted during network training and a cost function evaluated at each training epoch using available truth labels.
  • the artificial neural network (ANN) training is accomplished by a standard back-propagation algorithm using a cross-entropy cost function.
  • ANN artificial neural network
  • output layer of the artificial neural network (ANN) for a binary classifier is implemented as the softmax function with two-element vector [1, 0] for CAD positive and [0, 1] for CAD negative case.
  • the features used as inputs to either an ANN classifier or a SVM classifier are selected from the following: length (time duration) of: S 1 , systole, S2, and diastole; ECG RR peak duration; amplitudes of SI, systole, S2, and diastole; ratios of corresponding amplitudes; and energy of spectral bands associated with SI, systole, S2 and diastole.
  • Other features such statistics (mean, variance, skewness and kurtosis) of the parameter distributions can also be used as features for input to either an ANN classifier or a SVM classifier.
  • a convolutional neural network is employed to analyze cross-correlation images of acoustic signals in order to provide for automating the process of acoustic feature extraction.
  • the network input consists of the three- dimensional array of cross-correlation data recorded from the left or right positions on the thorax 20 of the test-subject 22, and combined into a single array, for example, representing three individual channels associated with three different auscultatory sound sensors 12.
  • the convolutional neural network comprises several convolutional layers that each use a 5x5 kernel array to scan input data to build a structure of acoustic features with increasing complexity. Neuron activation uses rectified linear unit (ReLU) nonlinearity to produce output data.
  • ReLU rectified linear unit
  • the final stage of the network classifier is a fully connected neural net with an additional clinical information (age, gender, blood pressure, etc.) merged with the acoustic features identified by CNN.
  • the network output is a two-node layer implementing binary classification via softmax function applied to the incoming data.
  • Reported performance metrics include prediction accuracy, sensitivity, specificity, negative prediction value and positive prediction value. New patient data are classified by passing through the same pre-processing stages and feeding the computed cross-correlation image to CNN classifier.
  • Convolutional neural networks have proved to be very efficient for prediction and classification problems especially with large scale problems involving images.
  • the typical size of input image data can be quite large, which makes application of standard feed forward networks either impractical or even impossible due to huge number of parameters to be trained.
  • Convolutional neural networks accommodate the size of the problem by weight sharing within small number of neurons comprising a receptive field that is scanned over the 2-D input.
  • One benefit of using a convolutional neural network (CNN) for machine learning problems to the ability thereof to leam important features directly from data, so as to provide for bypassing the feature extraction stage that is used by support vector machine (SVM) and feed-forward artificial neural network (ANN) classifiers.
  • SVM support vector machine
  • ANN feed-forward artificial neural network
  • a typical convolutional neural network (CNN structure consists of one or more convolution and pooling layers that build a hierarchical structure of features with increasing complexity. Following convolution and max pooling, the extracted features are fed to a fully connected network at the final stage of convolutional neural network (CNN) classifier.
  • CNN convolutional neural network
  • FIG. 44 illustrates a convolutional neural network (CNN) architecture incorporating a single convolution layer. Each cell of the max pool contains the maximum value of a corresponding array of cells in the corresponding convolution layer.
  • the receptive field is a relatively small 2-D array of neurons (for example, 5 x 5) that is scanned across the input image while performing an associated cross-correlation operation.
  • a relatively small number of connected neurons provides for a relatively small number of corresponding weights to be adjusted.
  • the max polling operation provides for reducing the size of the input to the associated fully-connected neural network by selecting pixels with maximum intensity from the associated convolution layer. Similar convolution and max polling operations can be performed multiple times to extract the most significant features before submitting to an associated fully -connected neural network for classification.
  • CNN convolutional neural network
  • the convolutional neural network (CNN) classifier can be applied either directly to the auscultatory sound signals 16 (or filtered versions thereof), or to corresponding 2-D images generated therefrom, for example, using either a continuous wavelet transform or an associated decomposition thereof by wavelet packet transformation (WPT) .
  • WPT wavelet packet transformation
  • the coefficients of J lh level of decomposition can be transformed into a matrix with dimensions (N/2 J ) x 2 J . where N is the size of the time domain signal.
  • Such 2-D data can be used train the convolutional neural network (CNN) classifier in order to find any patterns associated with CAD.
  • the convolutional neural network (CNN) design includes specification of the number of convolution layers, the size of receptive fields (kernel size), number of channels processed simultaneously, filter properties, regularization. After finalization of its design, the convolutional neural network (CNN) can be trained using a training data set and then evaluated using an unseen test data set.
  • the open- source TensorFlow flexible deep learning toolkit and API - which provide for building high- performance neural networks for variety of applications - have been used to design and train the convolutional neural network (CNN) for detecting CAD.
  • CNN convolutional neural network
  • a second embodiment of a convolutional neural network incorporates a plurality of convolution stages (Conv-1, Conv-2, Conv-3, Conv-4 and Conv-5) that are interspersed with corresponding associated MaxPool stages, wherein the output of the final max pooling stage is fed into a first fully -connected neural network (FC1) with 128 neurons, the output of which is fed into a second fully-connected neural network (FC2) with 64 neurons, the output of which is processed by a SoftMax output operation given by:
  • the input to the convolutional neural network is a series of either FIGS, 41/42-sytle convolution images, or a series of FIG. 43-sytle CWT transform images, with one image for each of the six different auscultatory sound sensor 12, with each image, for example, given by a 100 by 200 array of pixels.
  • Each of the convolution stages (Conv-1, Conv-2, Conv-3, Conv-4 and Conv-5) provides for convolving the array of input elements (or pixels) with an either 3x3 or 5x5 kernel or weight array, which is scanned across the input image with a stride if one pixel. If zero padding is used, a single kernel generates one feature map with the size as the original image, but 2D convolution applied as follows:
  • each convolution layer has multiple kernels that are applied to the input image.
  • the max pooling layer is applied which is a simple mask 2x2 scanned across feature maps. At each position only one pixel with highest magnitude is selected which is essentially a decimation operation that reduces the size of feature map by 2 while selecting brightest pixels.
  • the CNN, ANN and SVM classifiers each provide an independent means of classifying the associated input features in order to determine whether or not CAD is likely.
  • These are alternative classification algorithms and a decision as to which one to use is typically based on their respective performance, for example, prediction accuracy.
  • the different algorithms could be combined to make a single prediction, for example, using either a voting scheme or a weighted sum of likelihoods.
  • the auscultatory coronary-artery-disease detection system 10 can present various views of the acoustic data (both unprocessed and processed) that was captured during the test for review by a clinician.
  • the clinician may strengthen his case for a particular diagnosis, either in agreement with or disagreement with the result produced by the system. Additionally, particular visualizations may help reassure the patient that the diagnosis is the correct one.
  • the system presents heart/arteries view comprising a graphical representation of the heart and coronary artery tree, highlighting the data in accordance with a Textual View.
  • the clinician is presented with the option to highlight the blockage in either: a Stacked Heartbeat View; a Bruit Identification View (either mode); a Bruit Analysis View; or a Receiver Operating Characteristic Curve.
  • FIG. 45 illustrates test results showing severity of obstructions and zonal location of each obstruction within coronary artery tree.
  • the tree diagram also indicates which arteries are critical, and which are not, with associated color coding in accordance with the amount of occlusion, for example, with the illustrated region of 50 occlusion highlighted in a reddish hue, and the region of 30% occlusion highlighted in a yellowish hue.
  • the system presents, in a textual manner, the system’s interpretation of the acoustic data, including: whether or not the patient has a clinical level of CAD; the count of obstructions detected, for each obstruction: the location (zone) of the obstruction; the percentage occlusion of the obstruction; or the type of blockage (soft plaque, hard plaque).
  • the system provides for presenting an associated confidence level, which indicates how confident the system is of each specific element of information presented, or, based on the patient’s demographic data (age, sex, BMI, medical and family history), the percentage of other patients who have a similar count, severity, and position of obstructions. From this view, the clinician may switch to any other view listed in this document for more information.
  • the system presents a Receiver Operating Characteristic (ROC) curve to highlight the system’s CAD detection algorithm sensitivity as a function of its false positive rate (1 -specificity).
  • the system will plot the ROC curve and calculate the Area Under the Curve (AUC) based on the patient’s demographics, and the system’s clinical trial results.
  • AUC Area Under the Curve
  • the coordinate that corresponds to the current positivity criterion will be highlighted.
  • the clinician will be able to display a modified graph if he commands the system to exclude specific parts of the patient’s demographic data.
  • the graph may or may not contain gridlines and/or data points defining the ROC curve, and it may or may not fill in the AUC.
  • the data may be visualized as a line graph (with or without underfill), where true positives, true negatives, false negatives, and false positives are displayed, with a line indicating the currently-used cutoff value.
  • the system provides for visualizing the acoustic data resulting from an analysis of the breath-held auscultatory sound signals 16.1 from the auscultatory sound sensors 12.
  • the system presents a heartbeat view comprising a graphical plot of the systolic and diastolic intervals of each heartbeat captured.
  • the horizontal axis of the graph represents captured heartbeats. This representation is a subset of the acoustic capture time, as some acoustic data acquired is discarded (poor quality, etc.) during the system’s analysis process. The duration of the SI and S2 sounds are also highlighted on this axis.
  • the vertical axis may comprise acoustic data captured from each of the system’s sensors - both acoustic and ECG.
  • a correlation procedure is performed to ensure that the data captured from each of the system’s sensors is aligned to one another.
  • the clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, and explore any discrepancies between the data captured by the ECG and acoustic sensors for any particular heartbeat.
  • FIG. 47 illustrates an example of the Heartbeat View wherein the acoustic view has been limited by the clinician to Sensor#3 only, and zoomed in to see only 2 heartbeats, with the ECG data also displayed.
  • the system presents a graphical plot of the systolic and diastolic intervals of each heartbeat captured.
  • the horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the systolic interval to the end of the diastolic interval.
  • the duration of the SI and S2 sounds are highlighted on this axis.
  • the vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture, where each heartbeat is itself a graph, with a vertical axis representing intensity.
  • the system highlights any unexpected acoustic signals captured, as such signals may be an indication of an obstruction or other cardiac condition. For example, in FIG. 48, unexpected acoustic signals are highlighted in red.
  • the clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
  • the system presents a graphical plot of unexpected acoustic signals captured during the diastolic interval by means of a line graph with underfill.
  • the horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the diastolic interval to the end of the diastolic interval.
  • the duration of the S2 sound is also highlighted on this axis.
  • the vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture, where each heartbeat is itself a graph, with a vertical axis representing intensity.
  • An underfill is used to visually highlight deviation from the baseline.
  • the clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
  • the system presents a graphical plot of unexpected acoustic signals captured during the diastolic interval by means of a spectrogram.
  • the horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the diastolic interval to the end of the diastolic interval.
  • the duration of the S2 sound is also highlighted on this axis.
  • the vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture.
  • the system For each diastole displayed, the system highlights any unexpected acoustic signals captured, as such signals may be an indication of an obstruction or other cardiac condition. Such highlighted areas indicate the intensity of the unexpected signal. Highlights are in the form of color, which indicate varying intensity of the signal.
  • the clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
  • the system presents a graphical time/frequency plot of unexpected acoustic signals captured.
  • the horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the systolic interval to the end of the diastolic interval. The duration of the SI and S2 sounds are highlighted on this axis.
  • the vertical axis represents the frequency (in Hz or kHz) of any unexpected acoustic signals, averaged from all captured data.
  • the color on the graph represents the intensity of that signal.
  • the graph could be represented in monochrome as a contour plot.
  • the system provides a User Interface, and associated navigation, is designed for use on tablets and smartphones, and thus uses common touch-screen user interface paradigms. For example, two fingers moving apart from one another can be used to zoom in to any area on any graph, and two fingers moving together can be used to zoom out. A single finger can be used to highlight any areas on the horizontal axis, and the graph can be zoomed in to that highlighted area by touching a button.
  • Touching any area of the graph provides information to the user (either in a pop-up window or beside/below the graph) on the values of the horizontal and vertical axes at that point, as well as the“height” information of that point if available (e.g. in the Matrix View). For example, in Stacked Heartbeat View, touching on an individual heartbeat would cause the system to provide the heartbeat number, the maximum intensity of that heartbeat (and the time on the horizontal axis at which the maximum intensity occurs), and the time on the horizontal axis corresponding to the touch. In the case of the Matrix View, touching on any area of the graph would cause the system to provide the frequency, time, and intensity corresponding to the coordinate on the graph that was touched.
  • zoom can be accomplished through a common desktop/laptop user interface paradigm, such as dedicated zoom in/out buttons in the UI, or mouse wheel scrolling.
  • any sound feature that originates from the heart acts as a single source for all auscultatory sound sensors 12.
  • the characteristic sound that the heart makes, dub-blub acts like two sounds coming from two separate locations.
  • sound travels fast enough in the body that all auscultatory sound sensors 12 would effectively receive each sound individually at substantially the same time.
  • the location of the sound source can be triangulated from the locations of the auscultatory sound sensors 12 with the three largest signal strengths, weighted by the relative strengths of those signals, with the resulting calculated location of the sound source being relatively closer to auscultatory sound sensors 12 with relatively stronger signals than to auscultatory sound sensors 12 with relatively weaker signals.
  • the corresponding lengths Al, A2, A3 of corresponding edges El, E2, E3 of an associated tetrahedral solid 98 are assumed to be inversely related to the corresponding associated signal strength PI, P2, P3 of the auscultatory sound sensors 12.1, 12.2, 12.3 located at a corresponding vertices VI, V2, V3 of the tetrahedral solid 98 to which the corresponding edge El, E2, E3 is connected.
  • the lengths Al, A2, A3 of the edges El, E2, E3 are otherwise arbitrary, although relatively shorter lengths provide for less geometric dilution of precision (GDOP) of the ultimate determination of the lateral, (X, Y) location of the fourth vertex V4 than to relatively longer lengths.
  • the lateral (X, Y) location of the sound source 94 on the X-Y plane 96 being the same as the lateral (X, Y) location of the fourth vertex V4
  • the lateral (X, Y) location of the sound source 94 may be determined by solving the following system of three equations in three unknowns (X, Y, Z) for the location (X, Y, Z) of the fourth vertex V4:
  • the resulting lateral (X, Y) location of the sound source 94 may then be displayed, for example, as a location on a silhouette of a torso, or transformed to a corresponding location on the image of the heart illustrated in FIG. 45.
  • FIG.52 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Spectrogram mode, to confirm success of PCI.
  • FIG. 53 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Line Graph with Underfill mode, to confirm success of PCI.
  • the ability to compare the current test with a previous test is critical to the clinicians understanding of the progression of a particular cardiovascular condition (either worsening, or improved as in the case of a PCI procedure). This capability is relevant to the Stacked Heartbeat View, the Bruit Identification View (both modes), and the Bruit Analysis View.
  • the graphs may be rendered locally or remotely (server-based) or both, depending on the capabilities desired by the clinician and the organization to which he belongs. In most use cases (tablet, phone, desktop, or laptop), the graph rendering will be done locally, either through a web- browser (full or embedded) on the client, or through a graphics library optimized for each specific supported client platform.
  • the rendering may be done on the server side - graphs may be generated and exported to JPEG (or other similar) format so that they can be emailed or sent via instant message to interested parties.
  • any reference herein to the term“or” is intended to mean an“inclusive or” or what is also known as a“logical OR”, wherein when used as a logic statement, the expression“A or B” is true if either A or B is true, or if both A and B are true, and when used as a list of elements, the expression“A, B or C” is intended to include all combinations of the elements recited in the expression, for example, any of the elements selected from the group consisting of A, B, C, (A, B), (A, C), (B, C), and (A, B, C); and so on if additional elements are listed.
  • indefinite articles “a” or “an”, and the corresponding associated definite articles“the’ or“said”, are each intended to mean one or more unless otherwise stated, implied, or physically impossible.
  • expressions“at least one of A and B, etc.”,“at least one of A or B, etc.”, “selected from A and B, etc.” and“selected from A or B, etc.” are each intended to mean either any recited element individually or any combination of two or more elements, for example, any of the elements from the group consisting of“A”,“B”, and“A AND B together”, etc.

Abstract

An auscultatory sound signal from at least one auscultatory sound-or-vibration sensor is segmented into a plurality of associated heart cycle segments responsive to associated R-peak locations of an electrographic envelope signal representing an envelope response to an even power of an associated electrographic signal from an ECG sensor. A representation an envelope responsive to an even power of said auscultatory sound signal within said at least one heart cycle is locally modeled about at least a second peak to provide for locating the start of diastole of said at least one heart cycle.

Description

MEDICAL DECISION SUPPORT SYSTEM
CROSS-REFERENCE TO RELATED APPLICATIONS The instant application claims benefit of the following: U. S. Provisional Application Serial No. 62/838,270 filed on 24 April 2019, and U.S. Provisional Application Serial No. 62/838,296 filed on 24 April 2019, each of which is incorporated herein by reference in its entirety.
BRIEF DESCRIPTION OF THE DRAWINGS
In the accompanying drawings:
FIG. 1 illustrates a block diagram of a coronary-artery-disease detection system;
FIG.2 illustrates a first aspect of a data recording module and a first aspect of an associated docking system, in accordance with a first aspect of the coronary-artery-disease detection system illustrated in FIG. 1;
FIG. 3 illustrates a fragmentary view of a human thorax and associated prospective locations of auscultatory sound sensors at associated right, sternum and left, second, third, fourth and fifth, inter-costal spaces, left posterior locations at the second and third inter-costal spaces, and locations proximate to the heart apex;
FIG. 4 illustrates a second aspect of a data recording module and a second aspect of an associated docking system, in accordance with a second aspect of the coronary-artery-disease detection system illustrated in FIG. 1;
FIG. 5a illustrates an auscultatory sound sensor coupled to the skin of a test-subject, by bonding via associated adhesive layers or surfaces on both sides of an adhesive interface;
FIGS.5b and 5c each illustrate an auscultatory sound sensor that is detached, and therefore fully decoupled, from the skin of a test-subject, wherein FIG.5b illustrates the associated adhesive interface detached from the skin of the test-subject, and FIG. 5c illustrates the associated adhesive interface detached from the auscultatory sound sensor;
FIGS. 5d through 5g each illustrate an auscultatory sound sensor that is partially coupled to, but debonded from, the skin of a test-subject;
FIG. 6 illustrates a test-subject reclined on a surface, with their torso inclined while capturing auscultatory sound signals from a plurality of auscultatory sound sensors attached to the thorax of the test-subject;
FIG. 7 illustrates a flowchart of a first aspect of an associated auscultatory-sound-sensing process that incorporates a process for detecting a decoupling of the associated auscultatory sound sensors from the skin of the thorax of a test-subject being diagnosed for a prospective abnormal cardiovascular condition, wherein the decoupling-detection process occurs after each block of breath-held auscultatory sound time-series data is acquired, and is based upon scaled time-series data and responsive to an associated pre-determined debond-detection threshold; FIG. 8 illustrates a flowchart of a first aspect of a process for acquiring auscultatory sound signals from the associated auscultatory sound sensors coupled to the skin of the thorax of the test- subject being diagnosed for a prospective abnormal cardiovascular condition;
FIG. 9 illustrates a plurality of six blocks of breath-held, auscultatory-sound-sensor time- series data recorded from an auscultatory sound sensor coupled to the skin of the thorax of a test- subject being diagnosed for a prospective abnormal cardiovascular condition;
FIGS. 10a- lOf respectively illustrate a simulation of successively recorded blocks of breath-held, sensor time-series data illustrated in FIG. 9, each illustrated with an expanded time scale, wherein FIGS. 10a- lOe illustrates a condition for which the auscultatory sound sensor is coupled to the skin of the test-subject, and FIG. lOf illustrates a condition for which the auscultatory sound sensor is decoupled from the skin of the test-subject;
FIG. 11 illustrates a flowchart of a process for determining a scale factor used to scale auscultatory-sound-sensor time-series data, the latter of which is analyzed to detect whether or not the associated auscultatory sound sensor is decoupled from the skin of the test-subject, wherein the scale factor provides for directly determining if the associated auscultatory sound sensor is detached from the skin of the test-subject;
FIGS. 12a-12f respectively illustrate time-series of the absolute values of the corresponding time-series data illustrated in FIGS. 10a- lOf, further illustrating a division of the block of breath-held, sensor time-series data into a plurality of associated data segments, with each data segment of sufficient width to nominally include sound from a single heartbeat, and with the peak values in each data segment marked, wherein FIGS. 12a-12e illustrates a condition for which the auscultatory sound sensor is coupled to the skin of the test-subject, and FIG. 12f illustrates a condition for which the auscultatory sound sensor is decoupled from the skin of the test-subject;
FIG. 13 illustrates an accelerometer on the thorax of a test-subject during a respiration cycle of the test-subject;
FIG. 14 illustrates a breath-hold detection process;
FIG. 15 illustrates a flowchart of a first aspect of a process for detecting whether or not an auscultatory sound sensor is debonded from the skin of a test-subject;
FIG. 16 illustrates an organization of data from an auscultatory sound sensor recorded by an auscultatory coronary-artery-disease detection system from a test subject;
FIG. 17 illustrates a flowchart of a noise detection process;
FIG. 18 illustrates a flowchart of a process for generating a matched noise filter;
FIG. 19 illustrates a flowchart of a process for evaluating the noise content in a spectral signal of an auscultatory sound signal; FIG. 20 illustrates a flowchart of a process for logging results from the noise-evaluation process of FIG. 19;
FIG. 21 illustrates a block diagram of a process for preprocessing and screening auscultatory sound signals;
FIG. 22a illustrates a process for pre-processing auscultatory sound signals from auscultatory sound-or-vibration sensors;
FIG. 22b illustrates a process for pre-processing electrographic signals from an ECG sensor;
FIG. 23 illustrates a process for segmenting auscultatory sound signals from auscultatory sound-or-vibration sensors, by heart cycle based upon an electrographic signal from an ECG sensor, and by heat phase based upon the auscultatory sound signals;
FIG. 24 illustrates an auscultatory sound signal from an auscultatory sound-or-vibration sensor;
FIG. 25 illustrates a corresponding electrographic signal from an ECG sensor, in correspondence with the auscultatory sound signal illustrated in FIG. 24;
FIG. 26 illustrates a process for identifying heart-cycle boundaries in an electrographic signal from an ECG sensor;
FIG. 27 illustrates a first process for generating an envelope of a signal, which is called from the process illustrated in FIG. 26;
FIG. 28 illustrates an envelope, and associated peaks, of a portion of the electrographic signal illustrated in FIG. 25, generated in accordance with the process illustrated in FIGS. 26 and 27;
FIG. 29 illustrates a first aspect of a process for validating the peaks of an electrographic signal, which is called from the process illustrated in FIG. 26;
FIG. 30 illustrates a plot of the auscultatory sound signal illustrated in FIG. 24 together with plot of an envelope of the corresponding electrographic signal illustrated in FIGS. 25 and 28;
FIG. 31 illustrates a plot of the auscultatory sound signal illustrated in FIG. 24, but over a relatively greater total period of time, with the auscultatory sound signal segmented based upon the peaks of the envelope of the associated electrographic signal, and presented as a beat stack, each illustrating a corresponding associated heart beat;
FIG. 32 illustrates a second process for generating an envelope of a signal, which is called from the process illustrated in FIG. 23;
FIG. 33 illustrates a rectified auscultatory sound signal and an associate envelope thereof determined in accordance with the process illustrated in FIG. 32; FIG. 34 illustrates a plot of the auscultatory sound signal that is shown rectified in FIG. 33, together with plots of the associated quadratic models of the associated envelopes in proximity to the associated SI and S2 heart sounds, illustrating time-points associated with zero-amplitude roots of the associated quadratic models that are used to locate associated heart phases;
FIG. 35 illustrates the auscultatory sound signal beat stack illustrated in FIG. 31, together with indicia showing the locations of the roots of the quadratic models associated with the SI and S2 heart sounds that are used to locate associated heart phases, with the auscultatory sound signals of the beat stack aligned with one another based upon the mean values of the roots of the S2 heart sounds;
FIG. 36 illustrates a process for identifying outliers in a diastole region of an auscultatory sound signal;
FIG. 37 illustrates a flow chart of a process for selecting valid heart cycles and analyzing the results therefrom;
FIG. 38 illustrates a portion an auscultatory sound signal during diastole, and also illustrates associated noise power thresholds;
FIG. 39a illustrates an auscultatory sound signal for a plurality of heart cycles;
FIG.39b illustrates the auscultatory sound signal for the plurality of heart cycles illustrated in FIG. 39a, with the heart cycles aligned with respect to the mean times of the S2 heart sound;
FIG. 40a illustrates the auscultatory sound signal for a plurality of heart cycles illustrated in FIG. 39b, including indications of the start of the S2 heart sound and the end of diastole;
FIG. 40b illustrates portions of the auscultatory sound signal during diastole for the plurality of heart cycles illustrated in FIG. 40a, with the heart cycles temporally normalized and resampled;
FIG. 41 illustrates a cross-correlation process;
FIG. 42 illustrates a image of a plurality of localized cross-correlation signals associated with auscultatory sound signals from a plurality of heart cycles;
FIG. 43 illustrates a time-frequency analysis based on a continuous wavelet transform of an auscultatory sound signal during diastole;
FIG. 44 illustrates a process for detecting coronary artery disease from features of cross correlation images generated from analysis of auscultatory sound signals during diastole for a plurality of heart cycles, incorporating a Convolution Neural Network (CNN) classifier with a single convolution layer;
FIG. 45 illustrates a display of results in a heart/arteries view and textual view;
FIG. 46a illustrates a displace of a Receiver Operating Characteristic (ROC) curve; FIG.46b illustrates line graph display of true positives, true negatives, false negatives, and false positives;
FIG.47 illustrates a display of a Heartbeat View comprising a graphical plot of the systolic and diastolic intervals of each heartbeat captured;
FIG. 48 illustrated a display of a Stacked Heartbeat View;
FIG. 49 illustrates a Bruit Identification View in a Line Graphs with Underfill mode,
FIG. 50 illustrates a display of a Bruit Identification View in a Spectrogram mode;
FIG. 51 illustrates a display of a Bruit Analysis View;
FIG. 52 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Spectrogram mode, to confirm success of PCI;
FIG. 53 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Line Graph with Underfill mode, to confirm success of PCI;
FIG. 54 illustrates a second aspect of a process for detecting R-peaks in an electrographic signal;
FIG. 55 illustrates an envelope, and associated peaks, of a portion of the electrographic signal that includes noise and other signal components to be ignored;
FIG.56 illustrates a second aspect of a process for validating the peaks of an electrographic signal, which is called from the process illustrated in FIG. 26, following the second aspect process for detecting R-peaks in an electrographic signal, illustrated in FIG. 26;
FIG. 57 illustrates a three-level Discrete Wavelet Transformation (DWT) process;
FIG. 58 illustrates the effective filter spectra associated with the outputs of the three-level Discrete Wavelet Transformation process illustrated in FIG. 57;
FIG. 59 illustrates a scaling function of a Daubechies 4 (db4) wavelet family;
FIG. 60 illustrates a wavelet function of a Daubechies 4 (db4) wavelet family;
FIG. 61 illustrates a kernel discrete wavelet transformation process;
FIG. 62 illustrates a three-level Wavelet Packet Transformation (WPT) process;
FIG. 63 illustrates the effective filter spectra associated with the outputs of the three-level Discrete Wavelet Transformation process illustrated in FIG. 62;
FIG. 64 illustrates a high-pass filtered auscultatory sound signal for a single heart cycle from a high-pass filter having a 30 Hz cut-off frequency;
FIG. 65 illustrates a Wavelet Packet Transformation energy map for the high-pass filtered auscultatory sound signal illustrated in FIG. 64, with either decomposition levels;
FIG. 66 illustrates a selected basis of frequency bins from the Wavelet Packet Transformation energy map illustrated in FIG. 64, selected using Shannon entropy as a cost function; FIG. 67 illustrates a Support Vector Machine classifier trained to distinguish two-class data samples with respect to an associated highest-margin hyperplane;
FIG. 68 illustrates a feed-forward artificial neural network for binary classification;
FIG. 69 illustrates a geometric construction that provides for determining a projected transverse location of an acoustic source with respect to a plane containing three auscultatory sound sensors;
FIG. 70 illustrates various phases of a heart cycle in relation to an electrographic signal;
FIG. 71 illustrates the SI, S2, S3 and S4 phases of a heart cycle in relation to an electrographic signal; and
FIG. 72 illustrates a second embodiment of a Convolution Neural Network (CNN).
DESCRIPTION OF EMBODIMENT(S)
Referring to FIGS. 1 and 2, an auscultatory coronary-artery-disease detection system 10 incorporates at least one auscultatory sound sensor 12 that is operatively coupled to a recording module 13 running at least a first portion 14.1 of a Data Recording Application (DRA) 14, 14.1 on a first specialized computer or electronic system comprising a first computer processor or FPGA (Field Programmable Gate Array) 15 and first memory 17 powered by a first power supply 19, which provides for recording and preprocessing auscultatory sound signals 16 from the at least one auscultatory sound sensor 12. For example, in one set of embodiments, the at least one auscultatory sound sensor 12 comprises a first group 12’ of three auscultatory sound sensors 12, 121’, 12r, 123’ physically interconnected end-to-end with one another, and physically and electrically interconnected with a first wireless transceiver 18; and a second group 12” of three auscultatory sound sensors 12, 121”, 122”, 123” physically interconnected end-to-end with one another, and physically and electrically interconnected with the first wireless transceiver 18, with both groups 12% 12” of auscultatory sound sensors placed on the skin of the thorax 20 of a test-subject 22, in acoustic communication therewith. Referring also to FIG. 3, the placement of the first group of auscultatory sound sensors 12’ in FIG. 1 is illustrated with the respective associated auscultatory sound sensors 12, 121’, 12r, 123’ in substantial alignment with the corresponding respective third R3, fourth R4 and fifth R5, inter-costal spaces on the right side 20R of the thorax 20, and the placement of the second group of auscultatory sound sensors 12” in FIG. 1 is illustrated with the respective associated auscultatory sound sensors 12, 121”, 122”, 123” in substantial alignment with the corresponding respective third L3, fourth L4 and fifth L5, inter-costal spaces on the left side 20L of the thorax 20. Prospective sensor locations R2-R5, S2-S5, and L2-L5 illustrated in FIG. 3 respectively refer to the second R2 through fifth R5 inter-costal spaces on the right side 20R of the thorax 20, the second S2 through fifth S5 inter-costal spaces at the center/sternum 20s of the thorax 20, and the second L2 through fifth L5 inter-costal spaces on the left side 20L of the thorax 20.
Furthermore, prospective left-side posterior sensor locations LP2 and LP3 illustrated in FIG. 3 respectively refer to at the second LP2 and third LP3 intercostal spaces locations on the posterior of the thorax 20. Yet further, prospective sensor locations HA-1 and HA-2 are proximate to the apex of the heart, either on the anterior or the posterior of the thorax 20. For example, in one set of embodiments, The auscultatory sound sensors 12, 121’, 12r, 123”, 121”, 122”, 123” are located at the second S2, L2, third S3, L3 and fourth S4, L4 inter-costal spaces at the sternum S2-S4 and left-side L2-L4 of the thorax 20.
As used herein, the term“auscultatory sound” is intended to mean a sound originating from inside a human or animal organism as a result of the biological functioning thereof, for example, as might be generated by action of the heart, lungs, other organs, or the associated vascular system; and is not intended to be limited to a particular range of frequencies— for example, not limited to a range of frequencies or sound intensities that would be audible to a human ear, - but could include frequencies above, below, and in the audible range, and sound intensities that are too faint to be audible to a human ear. Furthermore, the term“auscultatory-sound sensor” is intended to mean a sound sensor that provides for transducing auscultatory sounds into a corresponding electrical or optical signal that can be subsequently processed.
The auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” provide for transducing the associated sounds received thereby into corresponding auscultatory sound signals 16 that are preprocessed and recorded by an associated hardware-based signal conditioning/preprocessing and recording subsystem 25, then communicated to the first wireless transceiver 18, and then wirelessly transmitted thereby to an associated second wireless transceiver 26 of an associated wireless interface 26’ of an associated docking system 27, possibly running a second portion 14.2 of the Data Recording Application (DRA) 14, 14.2 on a corresponding second specialized computer or electronic system comprising an associated second computer processor or FPGA (Field Programmable Gate Array) 28 and second memory 30, both of which are powered by an associated second power supply 32, which together provide for recording and preprocessing the associated auscultatory sound signals 16 from the auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123”.
For example, in accordance with one set of embodiments, the hardware-based signal conditioning/preprocessing and recording subsystem 25 includes an amplifier— either of fixed or programmable gain,— a filter and an analog-to-digital converter (ADC). For example, in one set of embodiments, the analog-to-digital converter (ADC) is a 16- bit analog-to-digital converter (ADC) that converts a -2.25 to +2.25 volt input to a corresponding digital value of - 32, 768 to + 32, 767. Furthermore, in accordance with one set of embodiments of the amplifier, the amplifier gain is programmable to one of sixteen different levels respectively identified as levels 0 to 15, with corresponding, respective gain values of 88, 249, 411, 571, 733, 894, 1055, 1216, 1382, 1543, 1705, 1865, 2027, 2188, 2350 and 2510, respectively for one set of embodiments. In accordance with another set of embodiments of the amplifier, the amplifier gain is fixed at the lowest above value, i.e., for this example, 88. so as to provide for avoiding the relative degradation of the associated signal-to-noise ratio (SNR) that naturally occurs with the relatively high gain levels of the programmable-gain set of embodiments.
It should be understood that the associated processes of the Data Recording Application (DRA) 14 could be implemented either in software-controlled hardware, hardware, or a combination of the two.
For example, in one set of embodiments, either or both the recording module 13 or docking system 27 may be constructed and operated in accordance with the disclosure of U.S. Provisional Application No. 62/575,364 filed on 20 October 2017, entitled CORONARY ARTERY DISEASE DETECTION SYSTEM, or International Application No. PCT/US2018/056832 filed on 22 October 2018, entitled CORONARY ARTERY DISEASE DETECTION SIGNAL PROCESSING SYSTEM AND METHOD, each of which is incorporated by reference in its entirety. Furthermore, in accordance with one set of embodiments, the auscultatory coronary-artery-disease detection system 10 may further incorporate an ECG sensor 34, for example, in one set of embodiments, an ECG sensor 34’ comprising a pair of electrodes incorporated in a corresponding pair of auscultatory sound sensors 12, wherein the signal from the ECG sensor 34’ is also preprocessed and recorded by a corresponding different signal channel of the same hardware-based signal conditioning/preprocessing and recording subsystem 25 of the recording module 13 that is used to preprocess the signals from the one or more auscultatory sound sensors 12. Alternatively, the ECG sensor 34 may comprise a separate set of a pair or plurality of electrodes that are coupled to the skin of the test subject, for example, in one set of embodiments, a pair of signal electrodes 35, 35+/ in cooperation with a ground electrode 35°, wherein, referring to FIG. 3 (illustrating the locations of the electrodes 35, 35+/ , 35°), in accordance with one non-limiting embodiment, the signal electrodes 35, 35+/ span the heart from diametrically-opposed quadrants of the torso 44, and the ground electrode 35° is located in a different quadrant, orthogonally displaced from a midpoint of a baseline connecting the signal electrodes 35, 35+/ . Alternatively the signal electrodes 35, 35+/ could span across the top of the top of the thorax 20, or at any pair of locations commonly used for conventional ECG tests. Furthermore, in one set of embodiments, the recording module 13 and docking system 27 may each incorporate a corresponding respective USB interface 36.1, 36.2 to provide for transferring corresponding auscultatory sound signals 16 and or an electrographic signal 37 - from an associated ECG sensor 34, 34’— from the recording module 13 to the docking system 27, for example, rather than relying upon the first 18 and second 26 wireless transceivers of an associated wireless interface 26’. Further alternatively, either instead of, or in addition to, the wireless interface 26’ or the USB interface 36.1, 36.2, data may be transferred from the recording module 13 to the docking system 27 via a portable memory element, for example, either an SD memory card or a Micro SD memory card.
The functionality of the Data Recording Application (DRA) 14 is distributed across the recording module 13 and the docking system 27. For example, referring to FIG.2, in accordance with a first aspect 10’ of an auscultatory coronary-artery-disease detection system 10, 10’, the Data Recording Application (DRA) 14 spans across the recording module 13 and the docking system 27, with a first portion 14.1 comprising the hardware-based signal conditioning/preprocessing and recording subsystem 25 operative on the recording module 13, and a remaining second portion 14.2 operative on the docking system 27. Alternatively, as another example, referring to FIG. 4, in accordance with a second aspect 10” of an auscultatory coronary-artery-disease detection system 10, 10”, the Data Recording Application (DRA) 14 is operative entirely on the recording module 13.
The auscultatory sound sensor 12 provides for sensing sound signals that emanate from within the thorax 20 of the test-subject 22 responsive to the operation of the test-subject’s heart, and the resulting flow of blood through the arteries and veins, wherein an associated build-up of deposits therewithin can cause turbulence in the associated blood flow that can generate associated cardiovascular-condition-specific sounds, the latter of which can be detected by a sufficiently- sensitive auscultatory sound sensor 12 that is acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22. For some cardiovascular-conditions associated with, or predictive of, a cardiovascular disease, the sound level of these cardiovascular-condition-specific sounds can be below a level that would be detectable by a human using a conventional stethoscope. However, these sound levels are susceptible to detection by sufficiently sensitive auscultatory sound sensor 12 that is sufficiently acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22. For example, in one of embodiments, the auscultatory sound sensor 12 may be constructed in accordance with the teachings of U.S. Provisional Application No. 62/568,155 filed on 04 October 2017, entitled AUSCULTATORY SOUND SENSOR, or International Application No. PCT/US2018/054471 filed on 04 October 2018, entitled AUSCULTATORY SOUND-OR- VIBRATION SENSOR, each of which is incorporated by reference in its entirety. Furthermore, in another set of embodiments, the auscultatory sound sensor 12 may be constructed in accordance with the teachings of U.S. Pat. Nos. 6,050,950, 6,053,872 or 6,179,783, which are incorporated herein by reference. Referring also to FIG. 5a, the auscultatory sound sensors 12, 121’, 12 , 123’ , 121”, 122”, 123” are acoustically coupled to the skin 38 of the thorax 20 of the test-subject 22 via an acoustic interface 40, for example, via a hydrogel layer 40’, that also functions as an associated adhesive interface 42 that is attached to the associated auscultatory sound sensor 12, 121’, 12r, 123’ , 121”, 122”, 123” with a first adhesive layer 42.1, for example, either a first surface 40.1’ of the hydrogel layer 40’ or a first layer of double-sided tape 42.1’ on a first side of the acoustic/adhesive interface 40, 42, and that is attached to the skin 38 of the thorax 20 of the test- subject 22 with a second adhesive layer 42.2, for example, either a second surface 40.2’ of the hydrogel layer 40’ or a second layer of double-sided tape 42.2’ on the opposing, second side of the acoustic/adhesive interface 40, 42. When fully coupled - as illustrated in FIG. 5a,— the auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” is fully attached to the acoustic/adhesive interface 40, 42 via the first adhesive layer 42.1, 42.1% 40.1% and the acoustic/adhesive interface 40, 42 is fully attached to the skin 38 of the thorax 20 of the test- subject 22 via the second adhesive layer 42.2, 42.2% 40.2% so that sound signals from within the thorax 20 of the test-subject 22 can propagate otherwise unimpeded to the auscultatory sound sensor 12, 121’, 122’, 123’ , 12'", 122”, 123”, thereby providing for a maximum achievable level of the corresponding associated auscultatory sound signals 16, and thereby improving the prospect of detecting an associated abnormal cardiovascular condition - if present - therefrom. Referring to FIGS. 5b and 5c - with the acoustic/adhesive interface 40, 42 respectively detached from the skin 38 or detached from the auscultatory sound sensor 12, respectively,— if the auscultatory sound sensor 12, 12r, 12 , 123’ , 121”, 122”, 123” is completely detached from the skin 38 of the thorax 20 of the test-subject 22, and thereby fully decoupled therefrom, the resulting auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” would be substantially non- responsive to sound signals from within the thorax 20 of the test-subject 22. Referring to FIGS. 5d to 5g, if the auscultatory sound sensor 12, 12r, 122’, 123’ , 121”, 122”, 123” is partially attached to the skin 38 of the thorax 20 of the test-subject 22, and thereby partially decoupled therefrom - i.e., in a condition referred to herein as being debonded therefrom— the resulting auscultatory sound sensor 12, 121’, 12 , 123’ , 121”, 122”, 123” would be only partially responsive to sound signals from within the thorax 20 of the test-subject 22, but not sufficiently responsive to provide for an associated auscultatory sound signal 16 of sufficient amplitude to provide for reliably detecting a prospective associated abnormal cardiovascular condition. More particularly, FIGS. 5d and 5e respectively illustrate an acoustic/adhesive interface 40, 42 partially detached from skin 38, and an acoustic/adhesive interface 40, 42 partially detached from an auscultatory sound sensor 12, respectively. Furthermore, FIG. 5f illustrates an auscultatory sound sensor 12 attached to a wrinkled acoustic/adhesive interface 40, 42, and FIG. 5g illustrates an acoustic/adhesive interface 40, 42 attached to wrinkled skin 38. In anticipation of prospective problems with nature of the attachment of the acoustic/adhesive interface 40, 42, the Data Recording Application (DRA) 14 is provided with a means - described more fully hereinbelow
— for detecting if one or more auscultatory sound sensors 12, 121’, 12 , 123’ , 121”, 122”, 123” is, or are, either detached or debonded from the skin 38 of the thorax 20 of the test-subject 22, so as to provide for excluding data from auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” that are either detached or debonded, from the skin 38 of the thorax 20 of the test-subject 22 from being used to diagnose a prospective abnormal cardiovascular condition.
Generally, the adhesive interface 42 could comprise either a hydrogel layer 40’, for example, P-DERM® Hydrogel; a silicone material, for example, a P-DERM® Silicone Gel Adhesive; an acrylic material, for example, a P-DERM® Acrylic Adhesive; a rubber material; a synthetic rubber material; a hydrocolloid material; or a double-sided tape, for example, with either rubber, acrylic or silicone adhesives.
Referring to FIG. 6, it has been found that the quality of the auscultatory sounds acquired from a test-subject 34 can be improved if the torso 44 of the test-subject 34 is inclined, for example, in one set of embodiments, at an inclination angle Q of about 30 degrees above horizontal— but generally, as close to upright (i.e. 0= 90 degrees) as can be accommodated by the associated adhesive interface(s) 42 of the associated auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123”— which imposes a transverse component of gravitational force on each of the auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123” that is resisted by the associated adhesive interface(s) 42.
Referring to FIGS. 7-15, in accordance with a first aspect 700, an auscultatory-soundsensing process 700 provides for first determining a scale factor SF from an initially-acquired block of auscultatory-sound-sensor time-series data S, and initially determining from the scale factor SF whether one or more of the auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123” is detached from the skin 38 of the thorax 20 of the test-subject 22, wherein when multiplied by the scale factor SF, the values of the associated auscultatory-sound-sensor time-series data S are nominally within a range that is a predetermined percentage of the dynamic range of the associated data acquisition system (for example, that provides 16- bit signed digital values). If there is no detachment, the first aspect 700, the auscultatory-sound-sensing process 700 provides for acquiring successive blocks of auscultatory-sound-sensor time-series data S while the test-subject 22 is holding their breath, and determining from each block of auscultatory- sound-sensor time-series data S - using an associated predetermined debond-detection threshold
- whether or not one or more auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123” is debonded from the skin 38 of the thorax 20 of the test-subject 22, or whether there is excessive noise in the auscultatory-sound-sensor time-series data S. The auscultatory-sensor time- series data S is rejected if excessive noise is detected, and the test is aborted if one or more auscultatory sound sensors 12, 121’, 12 , 123’ , 121”, 122”, 123” has become decoupled from the skin 38 of the thorax 20.
More particularly, referring to FIG. 7, the first aspect 700 of the auscultatory-soundsensing process 700 commences with step (702) by initializing a data-block counter i to a value of zero, and then, in step (704), acquiring a block of Ns contiguous samples of auscultatory- sound-sensor time-series data S in accordance with a first aspect 800 of a data acquisition process 800. This initially-acquired data is then used to determine the scale factor SF that is used to determine whether or not one or more auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” is/are detached from the skin 38 of the thorax 20 of the test-subject 22, and then subsequently to scale subsequent blocks of time-series data S. The initial block of auscultatory- sound-sensor time-series data S may be acquired either with, or without, the test-subject 22 holding their breath, but typically with the test-subject 22 allowed to breath normally— for their comfort and convenience. The number of samples Ns to be acquired is given by the product of an acquisition period Si in seconds, times a sampling frequency Fs in Hz. For example, in one set of embodiments, in step (704), the initially-acquired block of auscultatory-sound-sensor time- series data S typically contains 10 seconds of data, which at a sampling frequency Fs of 24 KHz, results in Ns = Si *Fs = 240,000 samples.
More particularly, referring to FIG. 8, the data acquisition process 800 commences with step (802) by pre-filtering the electronic signal from the associated auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” with an analog anti-aliasing filter, for example, an analog second-order band-pass filter having a pass band in the range of 3 Hz to 2.5 KHz, for which the upper-cutoff frequency is sufficiently below the sampling frequency (i.e. no more than half) so as to prevent high frequency components in the signal being sampled from appearing as low frequency components of the sampled signal, i.e. so as to prevent aliasing. Following step (802), if, in step (804), the test-subject 22 need not necessarily hold their breath - as is the case for the initially-acquired block of auscultatory-sound-sensor time-series data S,— then, in step (806), the pre-filtered auscultatory sound signal 16 is sampled at the sampling frequency Fs and converted to digital form by the associated analog-to-digital (ADC) converter. Then, from step (808), the auscultatory sound signal 16 continues to be sampled in step (806) until Ns samples of the block of auscultatory-sound-sensor time-series data S have been acquired, after which, in step (810), the Ns samples of auscultatory-sound-sensor time-series data S are returned to step (704) of the auscultatory-sound-sensing process 700. For example, FIGS. 9 and 10a each illustrate a first block of auscultatory-sound-sensor time-series data S of duration So that was recorded from one of the auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123”.
Referring again to FIG. 7, following step (704), in step (706), the scale factor SF is determined from the initially-acquired block of auscultatory-sound-sensor time-series data S, in accordance with an associated scale-factor-determination process 1100. More particularly, referring to FIGS. 11 and 12a, the scale-factor-determination process 1100 commences with step (1102), wherein the block of auscultatory-sound-sensor time-series data S is divided into a plurality of data segments 46, for example, each of the same data-segment duration SD that nominally spans a single heartbeat, for example, about one second. For example, in FIG. 12a, the So = 10 second block of auscultatory-sound-sensor time-series data S is divided into Nseg= 10 data segments 46, each of SD =one second duration. FIG. 12a illustrates a time series |S| containing the absolute values of the auscultatory-sound-sensor time-series data S illustrated in FIG. 10a. As indicated by X’s in FIG. 12a, for each data segment 46, m spanning the range k=kMiN(m) to k=kMAx(m) of the auscultatory-sound-sensor time-series data S(k), the corresponding maximum absolute value of the auscultatory-sound-sensor time-series data S(k) is determined, as given by:
Figure imgf000015_0001
Then, in step (1104), the median of these maximum values is determined, as given by
Figure imgf000015_0002
Finally, in step (1106), the scale factor SF is determined, as given by:
Figure imgf000015_0003
wherein DRo-p is the nominal zero-to-peak dynamic range of the auscultatory-sound-sensor time-series data S after scaling, i.e. after multiplying the acquired values by the scale factor SF. For example, in one set of embodiments, the nominal zero-to-peak dynamic range is set to be about 80 percent - more broadly, but not limiting, 80 percent plus or minus 15 percent— of the zero-to- peak range of the associated analog-to-digital converter - for example, in one set of embodiments, a 16-bit signed analog-to-digital converter— used to digitize the auscultatory-sound-sensor time- series data S in step (806). In one set of embodiments, the scale factor SF is integer-valued that, for an attached and bonded auscultatory sound sensor 12, 121’, 12r, 123’ , 121”, 122”, 123”, ranges in value between 1 and 28.
If one or more of the associated auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” is detached from the skin 38 of the thorax 20 of the test-subject 22, then the associated level of the auscultatory sound signals 16 will be low - for example, at a noise level - resulting in a relatively large associated scale factor SF from step (1106). Accordingly, if, in step (1108), the scale factor SF is in excess of an associated threshold SFMAX, then the Data Recording Application (DRA) 14 is aborted in step (1110), and the operator 48 is alerted that the one or more auscultatory sound sensors 12, 121’, 12 , 123’ , 121”, 122”, 123” is/are detached, so that this can be remedied. For example, in one set of embodiments, the value of the threshold SFMAX is 28 for the above-described fixed-gain embodiment, i.e. for which the associated amplifier has a fixed gain of 88, feeding a 16- bit analog-to-digital converter (ADC) that provides for converting a +/- 5 volt input signal to +/- 32,767. Otherwise, from step (1108), if the value of the scale factor SF does not exceed the associated threshold SFMAX, in step (1112), the scale factor SF is returned to step (706) for use in scaling subsequently -recorded breath-held auscultatory sound signals 16.1
Referring again to FIG. 7, following step (706), in step (708), the value of the data-block counter i is incremented, so as to point to the next block of auscultatory-sound-sensor time- series data S to be recorded. If, while this next block is recorded, the auscultatory sound sensors 12, 121’, 12 , 123’ , 121”, 122”, 123” remain attached and bonded to the skin 38 of the thorax 20 of the test-subject 22, and the associated breath-held auscultatory sound signals 16.1 are not corrupted by excessive noise, then this next block of auscultatory-sound-sensor time-series data S will then be subsequently used to detect a prospective abnormal cardiovascular condition therefrom. The auscultatory sound signals 16 used to detect prospective abnormal cardiovascular conditions are recorded while the test-subject 22 holds their breath, the latter to prevent the resulting cardiovascular-based auscultatory sound signals 16 from being overwhelmed by breathing-related sounds that are substantially louder than cardiovascular-based sounds, thereby providing for improving the associated signal-to-noise ratio of the cardiovascular- based auscultatory sound signals 16.
Accordingly, in step (710), a next block of Ns contiguous samples of auscultatory-soundsensor time-series data S is acquired over an acquisition period Si in accordance with a first aspect 800 of a data acquisition process 800, during which time the test-subject 22 is instructed to hold their breath. For example, in one set of embodiments, the nominal acquisition period Si is 10 seconds - but at least 5 seconds,— which, at a sampling frequency Fs of 24 KHz, results in Ns = Si *Fs = 240,000 samples.
More particularly, referring again to FIG. 8, the data acquisition process 800 commences with step (802) by pre-filtering the electronic signal from the associated auscultatory sound sensor 12, 121’, 122’, 123 , 121”, 122”, 123” with the above-described analog anti-aliasing filter. Then, from step (804), because breath-held data is to be acquired, in step (812), the test-subject 22 is instructed by the operator 48 to hold their breath. In step (814), if the operator 48 observes that the test-subject 22 is compliant in holding their breath, or, additionally or alternatively, if this is confirmed by a below-described breath-hold detection process 1400, then upon manual initiation by the operator 48, in step (816), a sample counter j is initialized to a value of one, and, in step (818), the next sample of pre-filtered auscultatory sound signal 16 is sampled at the sampling frequency Fs and converted to digital form by an associated analog-to-digital (ADC) converter. This process continues until either Ns = Si *Fs samples have been acquired, or the test- subject 22 resumes breathing. More particularly, in step (820), if one or more addition samples remain to be acquired, and if the operator 48 continues to observe that the test-subject 22 is holding their breath, or, additionally or alternatively, if this is confirmed by a below-described breath-hold detection process 1400, then, in step (822) the sample counter j is incremented, and the next sample is acquired in step (818). Otherwise, from step (820), if either all Ns = Si *Fs samples have been acquired, or if either the operator 48 observes that the test-subject 22 has resumed breathing, or, additionally or alternatively, if this is confirmed by a below-described breath-hold detection process 1400, or, if the test-subject 22 indicates by their own separate manual switch input that they will resume breathing, then, in step (824), the data acquisition process 800 terminates, and the block of breath-held auscultatory-sound-sensor time-series data S containing Ns =./ samples is returned to step (710). In one set of embodiments, if, following step (814), the test-subject 22 is not holding their breath, the pre-filtered auscultatory sound signals 16 are also separately-recorded while waiting for the test-subject 22 to hold their breath, or resume doing so. In practice, the auscultatory sound signals 16 typically continue to be recorded between breath-held segments, that latter of which are identified by associated recorded start and stop times with respect to the associated continuous recording.
Referring also to FIG. 13, alternatively, or additionally, in step (814), the auscultatory coronary-artery-disease detection system 10 may further incorporate an accelerometer 50 operatively coupled to the thorax 20 of the test-subject 22 to provide an associated acceleration signal responsive to the motion of the thorax 20. With the test-subject 22 lying on their back at an inclined angle, for example, at about 30 degrees above horizontal, for example, as illustrated in FIG. 4 of U.S. Provisional Application No. 62/568,155 filed on 04 October 2017, entitled AUSCULTATORY SOUND SENSOR, which has been incorporated by reference herein in its entirety, the associated acceleration signal - operatively coupled to recording module 13 and possibly transmitted to the docking system 27— may be twice integrated either in recording module 13 or the docking system 27 to generate a measure of the peak-to-peak displacement of the thorax 20, which if greater than a threshold would be indicative of breathing by the test- subject 22. More particularly, referring also to FIG. 14, for an auscultatory coronary-artery-disease detection system 10 incorporating an accelerometer 50 operatively coupled to the thorax 20 of the test-subject 22, an acceleration signal 52 therefrom may alternatively or additionally be processed by an associated breath-hold detection process 1400 to provide for automatically determining - for example, in step (814) of the data acquisition process 800 illustrated in FIG. 8 - whether or not the test-subject 22 is breathing, responsive to the determination, from the acceleration signal 52, of a peak-to-peak displacement of the thorax 20 of the test-subject 22. More particularly, beginning with, and in, step (1402), the respective previous/initial values of thorax displacement Yo and thorax velocity Vo. respectively, are each initialized to values of zero; a sample counter i is initialized to an initial value, for example, zero; the respective minimum YMIN and maximum YMAX values of thorax displacement are each set equal to the (initial) value of thorax displacement Yo ; the values of the sample counter IMIN and IMAX at which the corresponding minimum YMIN and maximum YMAX values of thorax displacement occur are set equal to the initial value of the sample counter i; and a BreathingFlag is set to indicate that the test-subject 22 is breathing. Then, in step (1404), the current sample of thorax acceleration A is acquired. Then, in step (1406), if the sample counter i is equal to the initial value, i.e. i == 0, then, in step (1408), the previous value of thorax acceleration Ao (i.e. initially, the initial value thereof) is set equal to the value of the current sample of thorax acceleration A, i.e. Ao == A. and then, in step (1410), the sample counter i is incremented, after which the breath- hold detection process 1400 repeats with step (1404).
Otherwise, from step (1406), if the current sample of thorax acceleration A is not the first sample, then, in step (1412), the current value of thorax velocity V is calculated by integrating the previous Ao and current A measurements of thorax acceleration, for example, using a trapezoidal rule, as follows:
Figure imgf000018_0001
wherein dt is the time period between samples, i.e. dt=l/Fs. Then, in step (1414), the current value of thorax displacement Y is calculated by integrating the above-calculated previous Vo and current V values of thorax velocity, for example, again using a trapezoidal rule, as follows:
Figure imgf000018_0002
Then, in step (1416), the respective previous values of thorax acceleration Ao, thorax displacement Yo and thorax velocity Vo are respectively set equal to the corresponding current values of thorax acceleration A, thorax velocity V and thorax displacement T, respectively, that will each be used in subsequent iterations of steps (1412) and (1414). Then, in step (1418), if the current value of thorax displacement Y is greater than then current maximum value of thorax displacement TMAX - for example, as would result during a phase of chest expansion by the test-subject 22,— then, in step (1420), the current maximum value of thorax displacement YMAX is set equal to the current value of thorax displacement Y and the corresponding value of the sample counter IMAX associated therewith is set to the current value of the sample counter i. Otherwise, from step (1418) - for example, as would result from other than a phase of chest expansion by the test-subject 22,— if, in step (1422), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMAX associated with the maximum value of thorax displacement YMAX is not equal to a threshold value A (the relevance of which is described more fully hereinbelow), then in step (1424), if the current value of thorax displacement Y is less than then current minimum value of thorax displacement YMIN - for example, as would result during a phase of chest contraction by the test- subject 22, - then, in step (1426), the current minimum value of thorax displacement YMIN is set equal to the current value of thorax displacement Y and the corresponding value of the sample counter IMIN associated therewith is set to the current value of the sample counter i. From either steps (1420) or (1426), in step (1428), if the amount by which the current maximum value of thorax displacement YMAX is greater the current minimum value of thorax displacement YMIN meets or exceeds a displacement threshold A YMAX. then, in step (1430), the BreathingFlag is set to indicate that the test-subject 22 is breathing, after which, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404). Similarly, from step (1428), if the displacement threshold AYMAX is not exceeded, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404). Further similarly, from step (1424) - for example, as would result from other than a phase of chest contraction by the test-subject 22,— if, in step (1432), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMIN associated with the minimum value of thorax displacement YMIN is not equal to the threshold value A, in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404).
If, from step (1432), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMIN associated with the minimum value of thorax displacement YMIN is equal to the threshold value A— following a minimum chest contraction of the test-subject 22, in anticipation of subsequent chest expansion, wherein the threshold value A is greater than or equal to one,— then, in step (1434), the peak-to-peak thorax displacement AY is calculated as the difference between the current maximum YMAX and minimum YMIN values of thorax displacement, and, in step (1436), the maximum value of thorax displacement YMAX is set equal to the current value of thorax displacement Y, and the value of the sample counter IMAX at which the corresponding maximum value of thorax displacement YMAX occurred is set equal to the current value of the sample counter i, in anticipation of subsequently increasing magnitudes of the current value of thorax displacement Y to be tracked in steps (1418) and (1420).
Similarly, if, from step (1422), the amount by which the current value of the sample counter i exceeds the value of the sample counter IMAX associated with the maximum value of thorax displacement YMAX is equal to the threshold value A— following a maximum chest expansion of the test-subject 22, in anticipation of subsequent chest contraction, wherein the threshold value A is greater than or equal to one,— then, in step (1438), the peak-to-peak thorax displacement AY is calculated as the difference between the current maximum YMAX and minimum YMIN values of thorax displacement, and, in step (1440), the minimum value of thorax displacement YMIN is set equal to the current value of thorax displacement Y, and the value of the sample counter IMIN at which the corresponding minimum value of thorax displacement YMIN occurred is set equal to the current value of the sample counter i, in anticipation of subsequently decreasing magnitudes of the current value of thorax displacement Y to be tracked in steps (1424) and (1426).
Accordingly, the threshold value A, provides for a delay to assure that a most-recent extremum of displacement has been reached, either the current maximum YMAX or minimum YMIN values of thorax displacement, before calculating the associated peak-to-peak thorax displacement AY.
From either steps (1436) or (1440), in step (1442), if the amount of the peak-to-peak thorax displacement AY calculated in steps (1434) or (1438), respectively, meets or exceeds the displacement threshold AYMAX, then, in step (1444), the BreathingFlag is set to indicate that the test-subject 22 is breathing. Otherwise, from step (1442), if the amount of the peak-to-peak thorax displacement AY calculated in steps (1434) or (1438), respectively, does not exceed the displacement threshold AYMAX, then, in step (1446), the BreathingFlag is reset to indicate that the test-subject 22 is not breathing. Following either step (1444) or (1446), in step (1410), the sample counter i is incremented, after which the breath-hold detection process 1400 repeats with step (1404).
Referring again to FIG. 7, in step (712), a corresponding block of scaled auscultatory- sound-sensor time-series data S is calculated by multiplying the acquired block of auscultatory- sound-sensor time-series data S by the scale factor SF, and in step (714), the scaled auscultatory-sound-sensor time-series data S is analyzed by an associated debond detection process 1500 to determine whether or not any of the auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” was debonded from skin 38 of the thorax 20 of the test-subject 22 during the associated data acquisition process 800.
More particularly, referring to FIG. 15, the debond detection process 1500 commences with step (1502) by initializing a sample counter j to a value of one, and initializing a threshold counter TC to a value of zero, wherein the threshold counter TC is a count of the number of contiguous successive samples for which the value of the scaled auscultatory-sound-sensor time-series data S is less than an associated predetermined debond-detection threshold DT. For example, in one set of embodiments, the debond-detection threshold DT is set to a value that is about 20% of the achievable maximum value of the output from the analog-to-digital converter (ADC). Then, in step (1504), if the absolute value of the sample of scaled auscultatory-soundsensor time-series data S, i.e. |S (j)|, is less than the predetermined debond-detection threshold DT (or, alternatively, if |SF*S(j)| < DT, thereby precluding a need to separately calculate and store scaled auscultatory-sound-sensor time-series data S), then in step (1506), the threshold counter TC is incremented, after which, in step (1508), if the value of the threshold counter TC does not exceed the number of samples in ND successive data segments 46, i.e. TC < ND *6D *FS, and in step (1510), if the sample counter j does not exceed the number of samples Ns in the block of scaled auscultatory-sound-sensor time-series data S, then, in step (1512), the sample counter j is incremented, and the process continues again with step (1504). Otherwise, from step (1504), if the absolute value of the current sample of scaled auscultatory-sound-sensor time- series data S, i.e. |S (j)|, is not less than the predetermined debond-detection threshold DT - indicating that the auscultatory sound sensor 12, 121’, 12 , 123’ , 121”, 122”, 123” is not debonded from the skin 38 of the thorax 20 of the test-subject 22, - then, in step (1514), the threshold counter TC is reset to a value of zero and the process continues with step (1510). Otherwise, from step (1510), if the sample counter j exceeds the number of samples Ns in the block of scaled auscultatory-sound-sensor time-series data S or auscultatory-sound-sensor time-series data S— indicating that the entire block of scaled auscultatory-sound-sensor time-series data S or auscultatory-sound-sensor time-series data S has been screened,— then the debond detection process 1500 is terminated with step (1516) by returning an indication to step (714) of the auscultatory-sound-sensing process 700 that the associated auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” is not debonded from the skin 38 of the thorax 20 of the test-subject 22. Otherwise, from step (1508), if the value of the threshold counter TC exceeds the number of samples in ND successive data segments 46, then the debond detection process 1500 is terminated with step (1518) by returning an indication to step (714) of the auscultatory-soundsensing process 700 that the associated auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” is debonded from the skin 38 of the thorax 20 of the test-subject 22. For example, in one set of embodiments, the value of ND is equal to 4, and the value of & is equal to 1 second.
Returning to FIG. 7, in step (716), if, from step (714), the debond detection process 1500 detected that the associated auscultatory sound sensor 12, 121’, 12r, 123’ , 121”, 122”, 123” was debonded while acquiring the block of auscultatory-sound-sensor time-series data S, then the Data Recording Application (DRA) 14 is aborted in step (718), and the operator 48 is alerted that one or more auscultatory sound sensors 12, 121’, 12r, 123’ , 121”, 122”, 123” are debonded, so that this can be remedied. Otherwise, if, from step (714), the debond detection process 1500 did not detect that the associated auscultatory sound sensor 12, 121’, 122’, 123’ , 121”, 122”, 123” was debonded while acquiring the block of auscultatory-sound-sensor time-series data S, then, in step (720), if sufficient noise-screened data has not been gathered - for example, in one set of embodiments, a total duration of at least 65 seconds of recorded data,— then the auscultatory- sound-sensing process 700 continues with step (708).
In step (724), an associated noise detection (i.e. noise-screening) process - operating on either the block of scaled auscultatory-sound-sensor time-series data S, or the block of auscultatory-sensor time-series data S, in parallel with the debond detection process 1500 - provides for detecting if the block of auscultatory-sound-sensor time-series data S has been corrupted by excessive noise, and if so, from step (726), that block of auscultatory-sound-sensor time-series data S is ignored, and the auscultatory-sound-sensing process 700 continues by repeating step (710) to acquire a new block of auscultatory-sound-sensor time-series data S. Otherwise, from step (726), if the block auscultatory-sound-sensor time-series data S has not been corrupted by excessive noise, the process continues with the above-described step (720).
From step (720), if sufficient noise-screened data has been gathered for which the associated one or more auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123” were not debonded from the skin 38 of the thorax 20 of the test-subject 22 - for example, in one set of embodiments, a total duration of at least 65 seconds of recorded data,— then, in step (722), at least the composite set of blocks of breath-held auscultatory-sound-sensor time-series data S acquired in step (710) are subsequently analyzed by an associated Data Analysis Application (DAA) 54 operative on the docking system 27 - as illustrated in FIGS. 2 and 3 - so as to provide for detecting an abnormal cardiovascular condition of the test-subject 22. In addition to recording the segments of breath-held data, alternatively, all data may be recorded and provided to the Data Analysis Application (DAA) 54, along with an associated index that provides for identifying the corresponding associated breath-held portions thereof for which the associated auscultatory sound sensors 12, 121’, 122’, 123’ , 121”, 122”, 123” were neither detached nor debonded from the skin 38 of the thorax 20 of the test-subject 22, nor corrupted by noise. FIGS. 9 and 10a- lOf illustrate a simulation of six blocks of breath-held auscultatory- sound-sensor time-series data S recorded in accordance with the first aspect 700 of auscultatory-sound-sensing process 700, with respective durations of <¾, <¾, <%, <¾, Ss, and <¾ during which time periods the test-subject 22 was holding their breath, separated by periods Di , 2h, , 2b, and 2b of normal breathing, wherein FIGS. lOa-lOe illustrate breath-held auscultatory sound signals 16.1, 16.1’ from a normally-bonded auscultatory sound sensor 12, 121’, 12r, 123’ , 121”, 122”, 123” as illustrated in FIG. 5a, and FIG. lOf illustrates a breath-held auscultatory sound signal 16.1, 16.1” from a debonded auscultatory sound sensor 12, 121’, 12r, 123’ , 121”, 122”, 123”, for example, as illustrated in any of FIGS. 5d-5g, for example, as might be caused by excessive hair between the adhesive interface 40 and the auscultatory sound sensor 12, poor placement of the auscultatory sound sensor 12 on the thorax 20 of the test- subject 22, poor angular orientation of the auscultatory sound sensor 12 relative to the surface of the skin 38, or wrinkled adhesive interface 40 between the auscultatory sound sensor 12 and the skin 38, For purposes of simplicity of illustration, FIGS. lOb-lOe are identical to FIG. 10a. However, it should be understood that typically the amplitude of the auscultatory sound signals 16, 16.1 varies from heartbeat to heartbeat, and from one breath-held segment to another.
Alternatively, one or more of the auscultatory-sound-sensing process 700, the data acquisition process 800, the scale-factor-determination process 1300, or the de-bond detection process 1500 could be implemented with corresponding alternative processes disclosed in U.S. Application Serial No. 16/136,015 filed on 19 September 2018 - with particular reference to FIGS. 16 through 22 - which is incorporated by reference herein in its entirety.
Referring to FIG. 16, in one set of embodiments, all the data is recorded throughout the duration of the test, including segments both with and without breathing, and a set of index pointers are used to identify locations of associated events, for example, index pointer arrays io[] and h[] to store the sample locations at the beginning and end of breath-held data segments of the corresponding sampled auscultatory sound data Sm[] from the mth auscultatory sound sensor 12, and later-used index pointer arrays isi[] and [] to store the sample locations of the SI sound at the beginning of each heart cycle, and the S2 sound at the beginning of diastole respectively, wherein in FIG. 16, the symbol“B” is used to indicate a period of breathing, the symbol“H” is used to indicate a period of breath-holding, and the symbol“D” is used to indicate a period of diastole. A status array Status [m, k] indicates the measurement status of the kth breath-held data segment of the mth auscultatory sound signal 16, i.e. the sampled auscultatory sound data Sm() from the mth auscultatory sound sensor 12. Accordingly, step (728)— that provides for ignoring data— may be implemented by setting the corresponding value of the status array Status(m, k) to a value of IGNORE. Referring to FIGS. 17-20, a noise detection process 1700 called from step (724) provides for determining whether or not a particular segment of breath-held auscultatory sound signal 16.1 is corrupted by excessive noise, and if so, provides for flagging that segment as being excessively noisy so as to be prospectively excluded from subsequent processing. Generally the noise detection process 1700 generates, in step (1714), frequency-domain noise filters FH[] responsive to cross-correlations of frequency spectra of pairs of adjacent auscultatory sound sensors 12. Accordingly, sensor-index pointers ml[p] and m2[p] provide for identifying the associated auscultatory sound sensors 12, ml[p], m2 [ p ] of each pair, the latter of which is identified by a pair pointer p. For each pair of adjacent auscultatory sound sensors 12 in a set of adjacent pairs— selected by sensor-index pointers ml[p] and m2[p] in steps (1706) and (1710), respectively, wherein the pair pointer p is initialized to 1 in step (1704) - the noise detection process 1700 generates, in step (1714), a frequency-domain noise filter FH[] by cross- correlating, in step (1802), the frequency spectra FSA[] and FSB[]— generated in steps (1708) and (1712)— of each associated breath-held auscultatory sound signal 16.1: SA[] and SB[], wherein the values of the frequency-domain noise filter FH[] are generated by normalizing the frequency spectra of the cross-correlation function to a range of 0 to 1 in step (1804), then subtracting these values from unity, and then setting resulting values that are less than a noise floor to the value of the noise floor in step (1806). Accordingly, when used to multiply the frequency spectra FSA[] and FSB[] in step (1904) as called from steps (1718) and (1720) for frequency spectra FSA[] and FSB[], respectively, the frequency-domain noise filter FH[] provides for attenuating the components of the breath-held auscultatory sound signal 16.1: SA[] and SB[] that are correlated with one another. Accordingly, the operation of step (1904) provides for a matched filtering process to accentuate the underlying noise in the associated breath-held auscultatory sound signal 16.1: SA[] or SB[] by attenuating frequency-component portions thereof that are correlated with a corresponding breath-held auscultatory sound signal 16.1: SB[] or SA[] from the other auscultatory sound sensor 12, ml[p], m21 p | of the pair. Furthermore, in step (1904), the product of the frequency-domain noise filter FH[] with either of the frequency spectra FSA[] or FSB[] is inverse Fourier transformed back to the corresponding time domain noise signal SN[] Then, in steps (1908) through (1918), a plurality of Nm-point short-time Fourier transform (STFT) arrays are generated, for example, with NFFT=1024, with each overlapped with respect on one another by a half width, i. e. NFFT 12, after which the associated average spectral power array FX[] in dB is calculated in step (1920). Then, the average powers PLOW, PMID, PHIGH in three respective frequency ranges 20 Hz to 200 Hz, 200 Hz to 800 Hz, and 800 Hz to 1,500 Hz, respectively, is calculated in steps (1922), (1924) and (1926), respectively, wherein each average power PLOW, PMID, PHIGH is compared with a corresponding threshold— for example, in one set of embodiments, -20 dB, -50 dB and -30 dB, respectively - in step (1928). The corresponding breath-held auscultatory sound signal 16.1: SA[] or SB[] is then flagged in step (1930) as being noisy if any of the associated noise power thresholds are exceeded. Referring to FIG.20, if one auscultatory sound sensor 12 exceeds the noise threshold in a given kth breath- held segment of breath-held sampled auscultatory sound data Sm[io[k]: ii [k] ] , that auscultatory sound sensor 12, m is flagged in step (2008) as being noisy so as to later be ignored for that kth breath-held segment. If a particular auscultatory sound sensor 12, m exceeds the noise threshold for more than one breath-held segment of breath-held sampled auscultatory sound data Sm[io[k]: ii[k]], then that auscultatory sound sensor 12, m is flagged in step (2016) as being bad. For each breath-held segment k of data, the process of steps (1706) through (1722) repeats for each of NPAIRS pairs of adjacent auscultatory sound sensors. For example, referring to FIG.3, in one set of embodiments, there are three pairs of adjacent auscultatory sound sensors, i.e. NPAIRS=3, as follows: for p=l, the auscultatory sound sensors 12 at the left and sternum second intercostal spaces L2, S2; for p=2, the auscultatory sound sensors 12 at the left and sternum third intercostal spaces L3, S3; and for p=3, the auscultatory sound sensors at the left and sternum fourth intercostal spaces L4, S4. The process of steps (1704) through (1726) is repeated until all the breath-held segments k of data have been processed.
More particularly, referring to FIG. 17, the noise detection process 1700 commences with step (1702) by initializing a breath-held-segment pointer A to a value of 1, so as to provide for pointing to the first breath-held segment of data. Generally, the breath-held-segment pointer k provides for pointing to the kth breath-held segment of data, of duration d A, extending between sample locations io[k] and ii[k], as illustrated in FIG. 16. Then, in step (1704), the pair pointer p is initialized to a value of 1, and a noise counter NNOISY is initialed to a value of 0. Then, in step (1706), the kth breath-held segment of breath-held sampled auscultatory sound data Sml[p][io[k]: ii[k]] is selected as the sampled auscultatory sound data SA[] of the first auscultatory sound sensor 12, ml[p] of the pair p, and in step (1708) the Fourier Transform of the sampled auscultatory sound data SA[] is calculated as FSA[] = FFT(SA[]). Similar, then, in step (1710), the kth breath-held segment of breath-held sampled auscultatory sound data Sm2|| ||io|k]: ii[k]] is selected as the sampled auscultatory sound data SB[] of the second auscultatory sound sensor 12, ni2[p] of the pair p, and in step (1712) the Fourier Transform of the sampled auscultatory sound data SA[] is calculated as FSB[] = FFT(SB[]).
Then, in step (1714), the frequency-domain noise filter FH[] generated by a matched- noise-filter generation process 1800 that - referring also to FIG. 18 - commences in step (1802) with the cross correlation of the frequency spectra FSA[], FSB[] of the sampled auscultatory sound data SA[], SB[] of the first ml[p] and second m2[p] auscultatory sound sensors 12 of the pair p, wherein the resulting cross-correlation is stored in array FH[] and then normalized to a range of 0 to 1 in step (1804), and then inverted in step (1806), wherein each element of the normalized array FH[] is subtracted from 1, and if the result is less than a noise floor, is set to the value of the noise floor NoiseFloor. Then, in step (1808), the resulting frequency-domain noise filter FH[] is returned and subsequently used in steps (1716) and (1718) of the noise detection process 1700 to evaluate the noise in the frequency spectra FSA[], FSB[] of the sampled auscultatory sound data SA[], SB[] of the first ml[p] and second m2[p] auscultatory sound sensors 12 of the pair p, respectively, in accordance with an associated noise-content- evaluation process 1900, the latter of which is called from step (1716) to evaluate the noise content of the frequency spectrum FSA[] of the sampled auscultatory sound data SA[] of the first auscultatory sound sensor 12, ml[p] of the pair p, and which is called from step (1718) to evaluate the noise content of the frequency spectrum FSB[] of the sampled auscultatory sound data SB[] of the second auscultatory sound sensor 12, m2[p] of the pair p.
Referring to FIG. 19, the noise-content-evaluation process 1900 commences in step (1902) with receipt of the index m of the associated auscultatory sound sensor 12, m, the associated frequency spectrum FS[] of the associated auscultatory sound sensor 12, m, and the associated frequency-domain noise filter FH[] Then, in step (1904), the time domain noise signal SN[] - containing a total of NPTS data points, i.e. the number of data points in the breath- held sampled auscultatory sound data Sm[io[k]: ii[k]] - is given by the inverse Fourier Transform of the product of the associated frequency spectrum FS[] with the associated frequency-domain noise filter FH[], corresponding to a time-domain cross-correlation of the corresponding associated time-domain signals. Then, in step (1906), each of the NFFT summation data points of a frequency-domain summation array FXSUM[] is initialized to zero, as is an associated summation counter NSUM, wherein the number of summation data points NFFT is a power of 2, for example, 1024, and substantially less than the total number of data points NPTS in the time domain noise signal SN[] Then, in step (1908), an index jMiN— to the first sample of an NFFT-point window of samples to be analyzed from the time domain noise signal SN[] - is initialized to a value of 1. Then, in step (1910), an index j\i \x— to the last sample of the NFFT- point window of samples to be analyzed from the time domain noise signal SN[] - is set to the value of jniiN + NFFT - 1. Then, in step (1912), if the end of the time domain noise signal SN[] has not been reached, then, in step (1914), the square of values - i.e. corresponding to noise power — of an NFFT-point Fourier Transform of the data from the NFFT-point window of samples of the time domain noise signal SN[] over the range of samples SN[JMIN] to SN[jMAx], is added to the frequency-domain summation array FXSUM[] Then, in step (1916), the summation counter NSUM is incremented, and in step (1918), the index jMiN is incremented by half the width of the NFFT-point window, i.e. by a value of NFFT 12, so as the provide for the next NFFT-point window to be analyzed to overlap the current NFFT-point window by half the window width. The above noise-content-evaluation process 1900 repeats beginning with step (1910), until, in step (1912), the index j MAX exceeds the end of the time domain noise signal SN[], after which, in step (1920), each of the values of the frequency-domain summation array FXSUM[] is divided by the value of the summation counter NSUM SO as to calculate an associated average noise power FX[] Then, in steps (1922), (1924) and (1926), respectively, the noise power is summed in three different corresponding respective frequency ranges, for example, 20 Hz to 200 Hz, 200 Hz to 800 Hz, and 800 Hz to 1,500 Hz, respectively, to give three corresponding respective power levels, PLOW, PMID and PHIGH, respectively. Then, in step (1928), if any of the power levels, PLOW, PMID or PHIGH exceeds a corresponding respective power threshold value ThresholdLOw, ThresholdMiD or ThresholdmGH, then, in step (1930), the noise threshold is considered to have been exceed for the particular auscultatory sound sensor 12, m and the particular associated segment k of breath-held sampled auscultatory sound data Sm[io[k]: ii[k]] is flagged with a corresponding indication of an associated status in an associated NoiseStatus flag. Otherwise, from step (1928), if none of the power levels, PLOW, PMID or PHIGH exceeds the corresponding respective power threshold value ThresholdLOw, ThresholdMiD or ThresholdmGH, then, in step (1932), the noise threshold is considered to have not been exceed for the particular auscultatory sound sensor 12, m and the particular associated segment of breath-held sampled auscultatory sound data Sm[io[k]: ii [k] ] is flagged with a corresponding indication of the associated status in the NoiseStatus flag. The results from either steps (1930) or (1932) are then logged by an associated results-logging process 2000, which is called from step (1934).
Referring to FIG. 20, the results-logging process 2000 commences with receipt in step (2002) of the index m of the associated auscultatory sound sensor 12, m and the associated NoiseStatus flag. If, in step (2004), the NoiseStatus flag indicates a noisy auscultatory sound sensor 12, m, then, in step (2006), the noise counter NNOISY is incremented, and, in step (2008), the corresponding element of the status array Status [m, k] for the associated auscultatory sound sensor 12, m and breath-held segment k is updated to activate an associate Noisy flag, so as to indicate the associated auscultatory sound sensor 12, m being noisy for the associated breath- held segment k. Then, in step (2010), if the value of the noise counter NNOISY is greater than 1, then, in step (2012), the corresponding elements of the status array Status [m, k] for each of the auscultatory sound sensors 12, m is updated to activate the Ignore flag, so as to provide for ignoring each of the auscultatory sound sensors 12, m for the associated breath-held segment k. Then, or otherwise from step (2010), in step (2014), if the Noisy flag of the status array Status [m, k] is activated for at least NFAIL breath-held segments k for the associated auscultatory sound sensor 12, m, then, in step (2016), status array Status [m] for the associated auscultatory sound sensor 12, m is updated to activate the Bad flag for the associated auscultatory sound sensor 12, m, so as to indicate that the associated auscultatory sound sensor 12, m is bad. Then, or otherwise from either step (2004) or step (2014), the results-logging process 2000 terminates by returning to the noise detection process 1700.
More particularly, referring again to FIG. 17, in step (1720), if all of the pairs of adjacent auscultatory sound sensors 12 have not been processed, then, in step (1722), the pair pointer p is incremented, and the noise detection process 1700 is repeated for the next pair of auscultatory sound sensors 12, ml[p], ni2[p] beginning with step (1706) for the same breath-held segment k. Otherwise, from step (1720), if in step (1724), additional segments of breath-held sampled auscultatory sound data Sm[io[k] : ii [k] ] remain to be processed, then, in step (1726), the breath- held-segment pointer k is incremented, and the noise detection process 1700 is repeated beginning with step (1704) for the next segment of breath-held sampled auscultatory sound data Sm[io[k] : ii[k]]. Otherwise, from step (1724), the noise detection process 1700 terminates with step (1728).
Referring again to FIG. 1, in accordance with one set of embodiments, the results from the docking system 27 may be transferred to a server computer system 56, for example, for subsequent transfer to an external data storage or processing system 58, for example, to provide for either viewing or analyzing the results at a remote location. Referring again to FIGS. 2 and 4, in one set of embodiments, the composite set of blocks of breath-held auscultatory-soundsensor time-series data S are screened prior to analysis by the associated Data Analysis Application (DAA) 54, for example, by first segmenting the set of blocks of breath-held auscultatory-sound-sensor time-series data S by heart cycle with an associated segmentation process 60, and then validating the associated heart cycle data with an associated heart cycle validation process 62, for example, to provide for additional screening for heart-phase associated noise.
Referring to FIG. 21, an auscultatory sound signal preprocessing and screening process 2100 provides for preprocessing breath-held auscultatory sound signals 16.1 and an associated electrographic signal 37 from an ECG sensor 34, 34’, for example, as acquired, for example, in accordance with the above-described auscultatory-sound-sensing process 700, with concurrent recording of an associated electrographic signal 37 from an ECG sensor 34, 34’. In accordance with one embodiment as illustrated, in step (2102) a good beat counter GB is initialized to zero, and a segment counter ksEG is initialized to a value of 1, wherein the good beat counter GB provides a count of the number of heart cycles for which the associated breath- held sampled auscultatory sound data S[] is not corrupted by either outliers or noise during diastole, and the segment counter ksEG is the value of the breath-held-segment pointer k for the current block B of breath-held sampled auscultatory sound data S[]
Then, from step (2104), referring to FIG. 22a, a segment of breath-held auscultatory sound signals 16.1 is acquired and preprocessed by an associated auscultatory sound signal acquisition and filtering process 2200, wherein, in step (2202), for each auscultatory sound sensor 12, in step (2204), a block of data of associated breath-held auscultatory sound signals 16.1 is acquired, for example, in accordance with the auscultatory-sound-sensing process 700, or as an associated segment k of a previously acquired and recorded breath-held sampled auscultatory sound data Sm[io[k]: ii[k] for the mth auscultatory sound sensor 12. With data recorded at a sampling frequency Fs of 24 kHz, and for portions of the spectrum of the breath- held auscultatory sound signals 16.1 of interest related to coronary artery disease limited to a few hundred Hertz - i.e. without significant acoustic energy above 500 Hz— the sampling rate may be reduced to 2 kHz. Accordingly, in step (2206), the breath-held auscultatory sound signal 16.1 is filtered by a fourth order Type II Chebyshev filter low-pass filter having a cut off frequency of 500 Hz to avoid aliasing, and then, in step (2208), decimated to a 2 kHz sampling rate using a phase invariant method that also involves low-pass filtering, so as to generate corresponding filtered-decimated breath-held sampled auscultatory sound signal 64.
The breath-held auscultatory sound signal 16.1 in some cases can include very low frequency - for example, around 10 Hz— vibrations that are believed to be associated with movements of the entire auscultatory sound sensor 12 stimulated by chest motion. Considering the sensor housing as an inertial mass under a tangential component of gravitational force attached to elastic surface, it is possible to initiate sensor vibration by small surface displacements. Such vibrations can be amplified by resonance characteristics of the tissue-sensor interface. Depending on the Q-factor of tissue-sensor system, vibrations may decay very slowly, extending well into diastolic interval of heart beat, contaminating the signal of interest with relatively large amplitude unwanted interference. The net effect of such interference is an unstable signal baseline and distortion of the actual underlying heart sounds. Potential sources of noise relevant to digitized acquisition of acoustic signals include: electric circuit thermal noise, quantization noise from the A/D converter, electro-magnetic interference, power line 60 Hz interference and acoustic noises relevant to human physiology and the environment where recording is done (ambient noise). Generally thermal noise power and A/D converter quantization noise are very low for the bandwidth of interest and may be significant only for the signals with amplitude in microvolt region. Furthermore, recording artifacts may significantly reduce visibility of the signal of interest since these artifacts may have relatively high amplitude and may overlap in spectral content. These artifacts are due to uncontrolled patient movements, signals related to respiration, and sensor vibrations produced by the associated oscillating mass of the auscultatory sound sensor 12 coupled to the elastic skin tissue surface (cardio seismographic waves). The latter type of artifact may be caused by the inertial mass of the sensor housing and can be relatively high in amplitude due to resonance properties of the sensor-tissue interface. Although the frequency of such vibrations is relatively low (around 10 Hz), the associated relatively high amplitude thereof can result in an unstable signal baseline, which complicates the detection of target signals.
However cardiac activity may also produce low frequency signals - for example, as a result of contraction of heart muscle, or as a result of valvular sounds— that may have valuable diagnostic information. In some situations, the spectrum of the artifacts may overlap with the acoustic spectrum of cardiac signals such as myocardium vibrations. Therefore, it can be beneficial to reduce baseline instability so as to provide for recording primarily acoustic signals originating from the cardiac cycle.
In accordance with one set of embodiments, these very low frequency artifacts may be rejected by using additional signal filtering to suppress the associated characteristic vibration frequencies, for example, using a software-implement high-pass filter 66 having a 3 dB cut-off frequency above 10 Hz, for example, as provided for by a Savitzky-Golay-based high-pass filter 66’, wherein, in step (2210), the filtered-decimated breath-held sampled auscultatory sound signal 66 is smoothed by a Savitzky-Golay (SG) smoothing filter 68 to generate a smoothed breath-held sampled auscultatory sound signal 70, the laher of which, in step (2212), is subtracted from the filtered-decimated breath-held sampled auscultatory sound signal 64 to then generate the corresponding resulting high-pass-filtered breath-held sampled auscultatory sound signal 72 having a relatively-higher signal-to-noise ratio (SNR), but without significant distortion of the original filtered-decimated breath-held sampled auscultatory sound signal 64
The digital Savitzky-Golay smoothing filter 68 is useful for stabilizing baseline wandering (for example, as may be exhibited in ECG signals) and provides for removing the low- frequency signal components without causing significant ringing artifacts. The Savitzky-Golay smoothing filter 68 employs a least squares approximation of a windowed signal using a polynomial function. The associated parameters of the Savitzky-Golay smoothing filter 68 include the window size M in samples that defines the associated cut-off frequency, and the polynomial degree N used for the approximation. The associated roll-off range is fairly wide and the cut-off frequency is somewhat arbitrary. For example, for the Savitzky-Golay smoothing filter 68 used in step (2210), the associated window size M - expressed in terms of window time duration tw - is in the range of 8 milliseconds to 100 milliseconds, with N=3. For example, a window time duration tw = 8 milliseconds provides a cut-off frequency of approximately 100 Hz, and a window time duration tw = 25 milliseconds, provides for passing signal frequencies above 40 Hz through the Savitzky-Golay-based high-pass filter 66’.
The Savitzky-Golay smoothing filter 68 is defined by a least-squares fit of windowed original samples, with an Nth degree polynomial,
Figure imgf000031_0001
so as to minimize the associated error function, E:
Figure imgf000031_0002
wherein the total window width is 2M+1 samples. The associated short-time window sliding through entire time series fits the data with a smooth curve. The frequency response of the Savitzky-Golay smoothing filter 68 depends strongly on the window size M and polynomial order N. The normalized effective cut-off frequency of the Savitzky-Golay smoothing filter 68 is empirically given as follows, wherein fc = OOs/p, for which oos.is the radian sampling frequency:
N + l
fc = (8)
3.2M - 4.6
Following the high-pass filter 66, 66’ of steps (2210) and (2212), from step (2214), the auscultatory sound signal acquisition and filtering process 2200 is repeated beginning with step (2202) for each of the auscultatory sound sensors 12, after which, from step (2216), the resulting blocks of high-pass-filtered breath-held sampled auscultatory sound signals 72— for each of the auscultatory sound sensors 12— are returned to step (2104).
Then, also from step (2104), referring to FIG. 22b, a corresponding segment of a concurrent electrograhic signal 37 is acquired and preprocessed by an associated electrographic signal acquisition and filtering process 2200’, wherein, in step (2204’), a block of electrographic data 74 of the corresponding electrographic signal 37 from the ECG sensor 34, 34’ - in correspondence with the blocks of high-pass-filtered breath-held sampled auscultatory sound signals 72 returned by the above-described auscultatory sound signal acquisition and filtering process 2200 - is acquired, for example, in accordance with the above-described data acquisition process 800, or as an associated segment k of previously acquired and recorded electrographic data 74. Then, in step (2206'), and as explained more fully herein below, the electrographic data 74 is filtered by a fourth order Type II Chebyshev filter low-pass filter having a cut-off frequency of 40 Hz, and then, in step (2208’), is decimated by a factor of ten, so as to generate corresponding filtered-decimated electrographic signal 76, which, in step (2216'), is returned to step (2104). Referring again to FIG. 21, in step (2106), the high-pass-filtered breath-held sampled auscultatory sound signals 72, or alternatively, the breath-held auscultatory sound signals 16.1, are segmented to identify the starting and ending points of each heart cycle, and to identify the starting and ending points of each associated diastole region or phase, thereof.
In accordance with a first aspect, this may be accomplished using the breath-held auscultatory sound signals 16.1— or the corresponding associated high-pass-filtered breath- held sampled auscultatory sound signals 72 - alone, without relying upon the associated electrograhic signal 37 from the ECG sensor 34, 34’, or upon the corresponding associated filtered-decimated electrographic signal 76, for example, in accordance with the following portions of the disclosure and drawings of U.S. Patent No. 9,364,184: Abstract, FIGS. 1-50, Col. 1, Line 1 through Col. 3, line 59 (indicated), Col. 5, line 1 through Col. 34, line 55 (indicated), and the claims, which are incorporated herein by reference.
However, in accordance with a second aspect, the electrographic signal 37 from the ECG sensor 34, 34’, and particularly, the corresponding associated filtered-decimated electrographic signal 76 responsive thereto, provides an effective basis for segmenting the breath-held auscultatory sound signals 16.1, 72 by heart cycle, after which the high-pass-filtered breath- held sampled auscultatory sound signal 72 may then be used to locate the associated dominant SI and S2 heart sounds that provide for locating the associated heart phases of systole and diastole, data from the latter of which provides for detecting coronary artery disease responsive to information in the breath-held auscultatory sound signals 16.1, 72.
Referring also to FIG. 23, in accordance with the second aspect, a heart-cycle segmentation and heart-phase identification process 2300 is called from step (2106) of the auscultatory sound signal preprocessing and screening process 2100 to locate the heart cycles responsive to the filtered-decimated electrographic signal 76, to then segment the high-pass- filtered breath-held sampled auscultatory sound signal 72 by heart cycle responsive thereto, and to then identify the region of diastole within each heart cycle (and implicitly, to therefor also identify the remaining region of systole) from an analysis of the high-pass-filtered breath-held sampled auscultatory sound signal 72 alone.
The normal human cardiac cycle consists of four major intervals associated with different phases of heart dynamics that generate associated audible sounds: 1) the first sound (SI) is produced by closing of mitral and tricuspid valves at the beginning of heart contraction, 2) during the following systolic interval, the heart contracts and pushes blood from ventricle to the rest of the body, 3) the second sound (S2) is produced by the closing of aortic and pulmonary valves and 4) during the following diastolic interval, the heart is relaxed and the ventricles are filled with oxygenated blood. Referring to FIGS. 24 and 25, respectively illustrating a breath-held auscultatory sound signal 16.1, 72 and a corresponding electrographic signal 37, 76 associated with six heartbeats, each QRS complex 78 illustrated in FIG. 25 is generated by an associated ventricular depolarization that precedes cardiac contraction. The SI heart sound resulting from the closing of mitral and tricuspid valves is observed immediately after R-peak , and the S2 heart sound resulting from the closing of aortic and pulmonary valves is observed at the end of T-wave of the ECG cycle. This relative timing of the heart sounds SI, S2 and the QRS complex 78 of the associated electrographic signal 37, 76 provides for using the electrographic signal 37 to locate the beginning and end of each heart cycle without relying upon the breath-held auscultatory sound signal 16.1, 72 to do so. For each heartbeat, the SI heart sound - marking the start of systole - follows shortly after the R-peak, and the S2 heart sound - marking the end of systole and the beginning of diastole - follows thereafter, with diastole continuing until the next R-peak.
Although the QRS complex 78 is the most prominent feature, the electrographic signal 37, 76 may be distorted by low frequency baseline wandering, motion artefacts and power line interference. To stabilize the baseline, a Savitzky-Golay-based high-pass filter 66’ - similar to that used in steps (2210/2212) when filtering the breath-held auscultatory sound signal 16.1 - may be used to cancel low-frequency drift, for example, prior to the subsequent low-pass filtering, in step (2206’), by the fourth order Type II Chebyshev filter low-pass filter, for example, having a 40 Hz cut-off frequency, and/or prior to decimation of the sampling by factor of 10 in step (2208’), which together provide for both reducing high frequency noise and emphasizing QRS complexes 78.
More particularly, referring again to FIG. 23, and also referring to FIG. 26, the heart- cycle segmentation and heart-phase identification process 2300 commences in step (2302) by calling an electrographic segmentation process 2600 that provides for identifying the locations of the R-peaks. More particularly, in step (2602), a block of electrographic data x[] is received, and, in step (2604), normalized so as to have a maximum absolute value of unity. Then, referring also to FIGS. 27 and 28, from step (2606), beginning with step (2702) of an associated first envelope generation process 2700, the block of filtered-normalized electrographic data x[] is used to generate an associated electrographic envelope waveform 80, Fs[] that emphasizes the R-peaks of the filtered-normalized electrographic data x[], wherein, for each value of index k — set in step (2702)— the corresponding value of the electrographic envelope waveform 80, Fs[k] is set in step (2704) responsive to a sum of values within a sliding window containing a subset of Nw points, as follows:
Figure imgf000033_0001
Equation (9) is similar to a Shannon energy function, but with the associated discrete signal values raised to the fourth power - rather than a second power— to emphasize difference between R-peaks 80’ and baseline noise. The value of the electrographic envelope waveform 80, Fs[] is calculated for each of the NPTS values of index k until, in step (2706), all points have been calculated, after which, in step (2708), the electrographic envelope waveform 80, Fs[] is returned to step (2606) of the electrographic segmentation process 2600.
Returning to FIG. 26, in step (2608), in accordance with a first aspect, the peaks of the electrographic envelope waveform 80, Fs[] are located, for example, by finding indices k for which the associated value of the electrographic envelope waveform 80, Fs[k] is a local maximum exceeding a fixed, predetermined threshold. For example, FIG. 28 illustrates the electrographic envelope waveform 80, Fs[] and associated R-peaks 80’ thereof, for a portion of a 10-second recording of a filtered-decimated electrographic signal 76.
The electrographic segmentation process 2600 for finding R-peaks is quite simple and stable when signal-to-noise ratio (SNR) of recording is sufficiently high. Otherwise, when the electrographic data 74 is excessively noisy, additional signal processing such as discrete wavelet decomposition may be used to further enhance R-peaks and facilitate QRS detection. Occasional relatively high amplitude transients due to patient movements or ambient interference may produce false peaks unrelated to QRS complex, and therefore complicate segmentation.
Referring to FIG. 29, in step (2610), the R-peaks 80’ of the electrographic envelope waveform 80, Fs[] are then validated by a peak validation process 2900, both with respect to temporal location - i.e. relative spacing— and magnitude - i.e. deviation from a median value,— so as to verify that the detected R-peaks 80’ correspond to R-peaks 80’ rather than noise. More particularly, in step (2902), for each of the detected R-peaks 80’, in step (2904), if the difference in temporal locations of each pair of R-peaks 80’ is less than a time threshold TMIN, or if, in step (2906), the difference between the magnitude P(tpEAK(i)) of each R-peak 80’ and the median value of the magnitudes of all R-peaks 80’ is greater than a magnitude threshold VMAX, then, in step (2908), a status indicator associated with the R-peak 80’ being tested is set to IGNORE, in order to prevent the associated R-peak 80’ from being used for subsequent segmentation of the breath-held auscultatory sound signal 16.1, 72. Otherwise, from step (2910), the peak validation process 2900 terminates in step (2912) and returns to step (2612) of the electrographic segmentation process 2600, which, in turn, returns the peak locations tpEAK[] to step (2302) of the heart-cycle segmentation and heart-phase identification process 2300, wherein valid R-peaks 80’ satisfy both of the following conditions:
Figure imgf000034_0001
and
Figure imgf000035_0001
wherein tpEAK(i) is the time of the 1th R-peak 80’ and R(ίrEAk(ΐ)) is the corresponding magnitude of the R-peak 80’.
Referring again to FIG. 26, alternatively, in accordance with a second aspect, the R-peaks 80’ of the electrographic envelope waveform 80, Fs[] are located in step (2608) by an associated R-peak detection process 5400 that provides for discriminating valid R-peaks 80’ from noise or other anomalies. Referring to FIGS. 54 and 55, the R-peak detection process 5400 determines the peak values and associated samples times of the electrographic envelope waveform 80, Fs[] within a plurality of samples window 81, each of window width tw, and each offset from one another with an associated hop period tHOP. More particularly, the R-peak detection process 5400 commences with step (5402) by setting a window start time tMiN to point to the beginning of the electrographic envelope waveform 80, Fs[], and by initializing a window counter Nw to a value of 1. Then, in step (5404), the maximum value of the electrographic envelope waveform 80, Fs[] is determined within the sample window 81 extending between sample times tvuN and tvHN + tw, and stored in element Pw[Nw] of window-peak array Pw[], wherein, for example, the window width tw is sufficiently wide to span a single heart cycle, for example, with tw equal to about 1.5 seconds. Then, in step (5406), the corresponding sample time at which the associated peak value occurred is stored as a corresponding element tpEAK_w[Nw] of an associated window- peak-time array tpEAK_w[] · Then, in step (5408), the window start time tMiN is incremented by the hop size tHOP, and, in step (5410), if the prospective end tvuN + tw of the prospective next sample window 81 is before the end tEND of the electrographic envelope waveform 80, Fs[], then, in step (5412), if the most-recently detected R-peak 80’ is unique, then, in step (5414), the window counter Nw is incremented. Then, or otherwise directly from step (5412), the R-peak detection process 5400 repeats beginning with step (5404). For example, in one set of embodiments, the hop size tHOP is set to half the shortest heart cycle, but no less than 0.3 second, or a corresponding fraction of the window width tw, although it should be understood that the particular values of the window width tw and hop size tHOP are not limiting. Although depending upon the window width tw and hop size ¾OR, different sample windows 81 may potentially span the same R-peak 80’, step (5412) provides for ensuring that each of the R-peaks 80’ in the window-peak array Pw[] are unique, and that the associated sample times in the window-peak array Pw[] are in monotonically increasing order.
Then, from step (5410) following R-peak detection within the last sample window 81, in step (5416), a minimum peak- threshold PMINLIM is set equal to about 60 percent of the median amplitude of the R-peaks 80’ within the window-peak array Pw[], and, in step (5418), a maximum peak-threshold PMAXLIM is set equal to twice the median amplitude of the R-peaks 80’ within the window-peak array Pw[] Then, in steps (5420) through (5430), R-peaks 80’ that are outside those limits are ignored, so as to provide for ignoring noise or other spurious signal components. More particularly, beginning in step (5420), a second window counter iw— that provides pointing to the above-detected unique R-peaks 80’— is initialized to a value of 1, and a peak counter NPEAK— that provides for counting the number of R-peak 80’ within the above- determined amplitude thresholds— is initialized to a value of 0. Then, in step (5422), if the value of the second window counter iw is less than the number of windows Nw determined above in steps (5402) through (5414), then, in step (5424), if the magnitude of the currently -pointed-to R- peaks 80’, i.e. Pw[iw], is greater than the minimum peak-threshold PMINLIM and less than the maximum peak-threshold PMAXLIM - indicating a potentially valid R-peak 80’ - then, in step (5226), the peak counter NPEAK is incremented, and the corresponding magnitude and time of the associated R-peak 80’, i.e. Pw[iw] and tpEAK_w[iw], are stored as respective values in in a corresponding peak array P [NPEAK] and peak-time array ΪREAK[NREAK] (alternatively, the window-peak array Pw[] and the window-peak-time array tpEAK_w[] could be reused in place for storing these values). For example, referring again to FIG. 55, step (5424) provides for ignoring occurrences of noise and a noise or spurious signal— the magnitudes of each of which are less than the minimum peak-threshold PMINLIM, -- and provides for ignoring the example of a false R-peak 80’ that resulted from electro-static discharge (ESD), and which exceeded the maximum peak-threshold PMAXLIM . Then, in step (5430), the second window counter iw is incremented to point to the next entry of the window-peak-time array tpEAK_w[], and the R-peak detection process 5400 repeats with step (5422). If, from step (5422), all entries of the have been processed, then, in step (5432), the process returns control to step (2608) of the electrographic segmentation process 2600.
Referring again to FIG. 26, following step (2608) and completion of the associated second aspect R-peak detection process 5400, a second aspect of a peak validation process 5600 provides for validating the R-peaks 80’ detected by the second aspect R-peak detection process 5400. More particularly, referring also to FIG. 56, in steps (5602) through (5608), the associated heart-cycle period DT[IPEAK] between each pair of temporally-adjacent R-peaks 80’ is calculated in step (5604), indexed by a second peak counter ΪREAK that is initialized to a value of 1 in step (5602) and incremented in step (5608) so long as, in step (5606), the value thereof is less then NPEAK-1 that had been determined in steps (5424) and (5426) of the second aspect R-peak detection process 5400. Then, in step (5610), a minimum heart-cycle-period-threshold
DTMINLIM is set to 0.3 seconds (corresponding to a heart rate of 200 BPM). Then, in step (5612), a corresponding heart-rate HR[] array is calculated from the heart-cycle period DT[] array, wherein each element of the former is given by the inverse of the corresponding element of the latter, after which the median heart-rate HRMEDIAN is determined a the median of the values of the elements of the heart-rate array HR[] Then, in step (5618) a status STATUS[] array to initially indicate that each R-peak 80’ associated with the heart-rate array HR[] is“OK”, and the second peak counter IPEAK that is initialized to a value of 1. Then, if, in step (5618), the heart-cycle period DT [IPEAK] is not greater than the minimum heart-cycle-period-threshold DTMINLIM, or, in following step (5620), the heart-rate HR[] is not both greater than 60 percent of the median heart-rate HRMEDIAN AND less than 140 percent of the median heart-rate HRMEDIAN, then, in step (5622) the STATUS [IPEAK] of the corresponding R-peak 80’ is set to “IGNORE”. Then, or otherwise from steps (5618) and (5620), if, in step (5624), all R-peaks 80’ have not been processed, then, in step (5626), the second peak counter IPEAK is incremented, and the second aspect peak validation process 5600 repeats with step (5618). Otherwise, from step (5628), the process returns control to step (2610) of the electrographic segmentation process 2600.
Referring again to FIG. 23, in step (2304), for each auscultatory sound sensor 12, in step (2306), the associated high-pass-filtered breath-held sampled auscultatory sound signal 72 thereof is segmented based upon the locations of the associated R-peaks 80’ in the corresponding filtered-decimated electrographic signal 76. For example, FIG. 30 illustrates a correspondence between the breath-held auscultatory sound signal 16.1, 72 and the corresponding electrographic envelope waveform 80, Fs[], wherein the locations of the R-peaks 80’ in the latter are used to segment the former by heart cycle. Furthermore, FIG. 31 illustrates a stack of segments of the high-pass-filtered breath-held sampled auscultatory sound signal 72, each segment corresponding to a different corresponding heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9 - each resulting from a single corresponding heart beat— within the corresponding continuous high-pass-filtered breath-held sampled auscultatory sound signal 72 illustrated in FIG. 30.
The exact location of the first SI and second S2 heart sounds produced by the respective closing of atrioventricular and semilunar valves is somewhat ambiguous, and can be particularly difficult to locate in situations when the associated breath-held auscultatory sound signals 16.1 are recorded in an environment with relatively high level ambient noise, or if there are associated heart murmurs resulting from turbulent blood flow. However, even under relatively-poor recording conditions, the first SI and second S2 heart sounds of the cardiac cycle remain the most prominent acoustic features. Referring again to FIG. 23, following segmentation of the high-pass-filtered breath-held sampled auscultatory sound signal 72 into corresponding resulting associated heart-cycle segments in step (2306) — each comprising one heart cycle 82, resulting from a single corresponding heart beat, - beginning with step (2308), for each of the heart cycles 82 located in step (2306), in step (2310) a Shannon energy envelope is generated from the corresponding associated portion of the high-pass-filtered breath-held sampled auscultatory sound signal 72 in accordance with a corresponding second envelope generation process 3200, which is similar to the above-described first envelope generation process 2700 except for the associated envelope function. More particularly, referring to FIG. 32, beginning with step (3202) of the second envelope generation process 3200, for the kth of NSAMPLES samples of a block of high-pass- filtered breath-held sampled auscultatory sound data 72, s[] of the corresponding associated heart cycles 82, a corresponding value of an associated acoustic envelope waveform 84, Es[k] is generated in step (3204), responsive to a sum of values within a short-time sliding window - with different windows overlapping one another— containing a subset of Nw points, as follows in accordance with a canonical Shannon energy function Es[k] that provides a measure of the energy of the underlying high-pass-filtered breath-held sampled auscultatory sound signal 72, s[] :
Figure imgf000038_0001
The value of the acoustic envelope waveform 84, Es[k] is calculated for each of the NSAMPLES values of index k until, in step (3206), all points have been calculated, after which, in step (3208), the acoustic envelope waveform 84, Es[k] is returned to step (2310) of the heart-cycle segmentation and heart-phase identification process 2300.
For envelope generation by each of the first 2700 and second 3200 envelope generation processes illustrated in FIGS. 27 and 32, the associated signal x[], s[], from which the corresponding envelope Fs[], Es[] is generated, is zero-padded at the beginning and at the end with (Nw - l)/2 zeros, assuming Nw is an odd integer, wherein the first data point in the padded signal x[], s| | would then begin at index l+(Nw - l)/2. After that, sliding window with size Nw and stride = 1 is used to compute windowed values of the associated envelope function, so as to provide for generating a time-series of the same length as the associated signal x[], s[], without a time shift relative thereto.
Referring again to FIG. 23, in step (2312), the acoustic envelope waveform 84, Es[k] generated in step (2310) is smoothed to reduce local fluctuations therein, for example using either a moving average filter, or a Savitzky-Golay smoothing filter 68, for example, with an associated window size of tw = 8.3 milliseconds. FIG. 33 illustrates an example of an acoustic envelope waveform 84, Es[] in correspondence with a rectified version 86 (i.e. containing an absolute value) of the associated high-pass-filtered breath-held sampled auscultatory sound data 72, s[] from which the acoustic envelope waveform 84, Es[k] was generated, wherein the respective envelope peaks 88S1, 88s2 associated with the corresponding SI and S2 heart sounds can be readily identified in the acoustic envelope waveform 84, Es[].
Referring again to FIG. 23, in step (2314), the locations of the envelope peaks 88S1, 88s2 associated with the corresponding SI and S2 heart sounds are identified using an adaptive threshold method that iteratively adjusts an associated threshold value down from a maximum value until the two most prominent envelope peaks 88S1, 88s2 emerge within the associated heart cycle 82, above the a particular threshold limit. Referring also to FIG. 34, the final position of envelope peaks 88S1, 88s2 associated with the corresponding SI and S2 heart sounds is found as the apex of a local quadratic model 90.1, 90.2 of the acoustic envelope waveform 84, Es[], as described more fully hereinbelow. Adaptive thresholding provides for accommodating substantial variation in the relative magnitude of the envelope peaks 88S1, 88s2 that might occur either from one heart beat to another, from one patient to another, or from one recording site to another.
Then, in step (2316), the locations of the envelope peaks 88S1, 88s2 associated with the corresponding SI and S2 heart sounds are validated using a normalized acoustic envelope waveform 84, Es[], i.e. normalized to a range of 0 and 1, and the associated local quadratic models 90.1, 90.2 thereof, in accordance with a minimum-time-spacing criteria used to remove or ignore spurious transient peaks unrelated to heart sounds, similar to equation 10 above that is associated with step (2904) of the above-described peak validation process 2900 used to validate the electrographic envelope waveform 80, Fs[].
Then, in step (2318), the acoustic envelope waveform 84, Es[] is searched relative to the associated indices ksi_PEAK, ks2_pn \i<— respectively associated with the corresponding respective envelope peaks 88S1, 88s2— to find adjacent data points therein,— i.e. having associated indices ksi-, ksi+, ks2-, ks2+— for which the corresponding values of the acoustic envelope waveform 84, Es(ksi-), Es(ksi+), Es(ks2-), Es(ks2+) are each about 5 percent down, i.e. 95 percent of, the corresponding values Es(ksi PEAK), Es(ks2_PEAK> of the associated envelope peaks 88S1, 88s2.
Then, in step (2320), respective local quadratic models ESl(k), 90.1’ and ES2(k), 90.2’ are fitted - for example, by least-squares approximation— to the three points associated with each of the corresponding respective envelope peaks 88S1, 88s2 as follows:
ES l (k) =Quadratic Fit ( {ksi-, Es(ksi-)} , {ksi_PEAK, Es(ksi_PEAK)} , {ksi+, Es(ksi+)} ) (13a) ES2(k) =Quadratic Fit ( (ks2-, Es(ks2-)}, {ks2_PEAK, Es(ks2_PEAK)} , {ks2+, Es(ks2+)} ) (13b) Then, referring again to FIG. 34, in step (2322), the respective pairs of roots (ksi_sTART, ksi END) and (ks2_START, ks2_END) of the local quadratic models ESl(k), 90.1’ and ES2(k), 90.2’ are solved, such that ESl(ksi sTART) = ESl(ksi END) = 0 and ES2(ks2_START) = ES2(ks2_END) = 0, Wherein ksi START < ksi_PEAK< ksi END and ks2_START < ks2_PEAK < ks2_END.
Referring to FIGS. 70 and 71, in addition to the SI and S2 sounds, S3. S4 or S5 sounds may be present in the heart cycles 82 for some test-subjects 22. The purpose of identifying SI, S2, S3, S4. and S5 regions is to help extract or derive useful features (such as Indicators that correlate highly with the existence of coronary artery disease (CAD)) from heart sounds. In turn, these features can be used in direct detection and machine learning for discriminating CAD patients from non-CAD patients. The prerequisite of many feature extractions is to correctly identify the SI and S2 sounds.
The S3, S4, and S5 sounds do not appear for every patient. But each may indicate one or more cardiac conditions, such as, CAD, Aortic Insufficiency, Aortic Stenosis, Luminal Irregularities, and Mitral Regurgitation. The regions within diastole during which the S3 and S5 sounds may occur are located relative to the diastasis region of diastole, which is a relative rest period of the heart during mid-to-late diastole, during which period the heart motion is minimal. The region during which the S4 sound may occur is located relative to the R-peak 80’ at the end of the heart cycle 82 and the beginning of the next heart cycle 82.
The starting point of the diastasis region is determined using what is referred to as the Weissler or Stuber formula for the period of delay DT— in milliseconds— from an R-peak 80’ to the starting point of the diastasis region, given by the following:
DTms = (ATR-R ms - 350) .03 + 350
wherein ATR-R is the time interval in milliseconds between R-peaks 80’. In one set of embodiments, this is approximately the ending point for the region most likely to include S3, i.e. the S3 region. Accordingly, the starting point for the S3 region is determined by advancing relative to the starting point of diastasis - or, equivalently, the end point of the S3 region— by a time interval ATs3. For example, in this set of embodiments, the time interval commences at about 100 milliseconds prior to the starting point of diastasis and ends at the starting point of diastasis. In another set of embodiments, the time interval of the S3 region is taken to extend from about 60 milliseconds prior, to 60 milliseconds after, the starting point of diastasis. The S3 swing is calculated by subdividing the S3 region of the associated breath-held auscultatory sound signal 16.1, for example, band-pass-filtered breath-held sampled auscultatory sound data s[] having a pass-band in the range of 20, 25, or 30 Hz to 40 or 50 Hz, , into a series of - i. e. one or more— time intervals, and calculating or determining one or more of the difference between the maximum and minimum amplitude values— i.e. maximum amplitude - minimum amplitude,— the minimum amplitude, and the maximum amplitude, for each interval in the S3 region.
In addition to the above time-domain analysis, the associated breath-held auscultatory sound signal 16.1 is also analyzed in the frequency domain using a Short-Time Fourier Transform (STFT), for example, in one set of embodiments, having a 1 Hz frequency resolution and a 0.0025 second time resolution— but generally, using frequency and time resolutions that may be empirically adjusted to improve detection and discrimination - in cooperation with an associate windowing method, for example, using Chebyshev windowing on a sliding window that is moved along the S3 region of the breath-held auscultatory sound signal 16.1. The frequency-domain features for each heart beat are generated by calculating the mean and median of each of the windows of the STFT.
The S3 swing values and frequency-domain features are saved for further calculations and/or for use as an input to one or more of the below-described classification processes. In view of there being relatively few patients that clearly exhibit an S3 sound, an unsupervised clustering method is applied on all the generated features to classify heart beats into two clusters that respectively include“with S3” and“without S3 heart beats. S3 is analyzed on a beat-by-beat basis. Given the S3 is not an intermittent signal, one strategy is to analyze all heart beats from a given patient, and if S3 appears in more than one third of all the heart beats, that patient would be identified as having S3 There are some patients who exhibit a low-ejection fraction ratio (e.g. an ejection fraction (E. F.) lower than 35%) that are highly likely to have S3. However some of these patients have CAD, and some have other types of cardiac arrhythmia. Once the unsupervised clustering is applied followed by voting among all beats belonging to one patient, if those patients are found in the cluster“with S3 this would provide for validating that the clustering matches reality.
The S4 region is located in a final portion of diastole, starting at the P wave on the ECG signal. However, in anticipation of a prospective absence of the P wave in some ECG recordings, an approximate starting point for the S4 region may be used, for example, a time interval of ATs4, for example, from 10 to 20 percent of the period of the heart cycle 82, for example, about 100 to 200 milliseconds in a 1 second duration heart cycle 82, in advance of the R-peak 80’ at the end of the heart cycle 82, or equivalently, in advance of the beginning of the SI sound of the next heart cycle 82. The S4 region of the associated breath-held auscultatory sound signal 16.1, e g. the band-pass-filtered breath-held sampled auscultatory sound data s[] having a pass-band in the range of 20, 25, or 30 Hz to 40 or 50 Hz, is subdivided into a series of intervals and the associated S4 swing within each interval is calculated as the absolute value of the difference between maximum and minimum amplitude magnitudes of the raw or high-pass-filtered data within that interval in the S4 region. The S4 swing is calculated separately for audible (above 20 Hz) and inaudible (below 20 Hz) frequency ranges, for which the configurations of the associated signal filters are specific to the particular frequency range. Generally, the minimum and maximum values of the signal will depend upon the associated frequency range of the associated signal, and the type of filter used to generate the filtered signal.
The S2 swing is also similarly calculated over the associated S2 region, in addition to calculating the S4 swing as described hereinabove. The ratio of the S4 swing to S2 swing provides a measure of the likelihood that a patient exhibits an S4 sound, with that likelihood increasing with increasing value of the S4-to-S2 swing ratio. For example, FIG. 71 illustrates a heartbeat for which the S4-to-S2 swing ratio has a value of 0.49 for the illustrated channel 1 of the band-pass- filtered breath-held sampled auscultatory sound data s[].
The S2 and S4 swing values, and/or the S4-to-S2 swing ratio, are saved for further calculations and/or for use as an input to one or more of the below-described classification processes. In accordance with one approach, the mean value of the S4-to-S2 swing ratio is calculated for an entire population of patients, beat by beat, and then the median of S4-to-S2 swing ratio is taken across all beats of each patient. For those patients for which the S4-to-S2 swing ratio is greater than the associated mean value of the S4-to-S2 swing ratio are identified as having an S4 sound for purposes of training the associated below-described classifier, from which an associated threshold value is determined that - in combination with other factors— provides for discriminating patients with CAD from patients without CAD, after which the threshold value of the S4-to-S2 swing ratio is then applied to the associated test set.
The S5 region is identified as the end of the S3 region to the start of the S4 region. Accordingly, the starting point is determined using the above-described Weissler or Stuber formula. As mentioned, this is approximately the ending point for the S3 region. The ending point of the S5 region is located at the beginning of a time interval of A TSJ. for example, about 100 milliseconds, in advance of the R-peak 80’ at the end of the heart cycle 82, or equivalently, in advance of the beginning of the SI sound of the next heart cycle 82. The S5 swing is calculated by subdividing the S5 region of the associated breath-held auscultatory sound signal 16.1, e.g. the band-pass-filtered breath-held sampled auscultatory sound data s[] having a pass-band in the range of 20, 25, or 30 Hz to 40 or 50 Hz, into a series of intervals and calculating the absolute value of maximum amplitude - minimum amplitude for each interval in the S5 region. The S5 swing values may be saved for further calculations and/or for use as an input to one or more of the below-described classification processes.
If, in step (2324), all heart cycles 82 in the high-pass-filtered breath-held sampled auscultatory sound signal 72 have not been processed, then the heart-cycle segmentation and heart-phase identification process 2300 is repeated beginning with step (2308) for the next heart cycle 82. Otherwise, from step (2326), if all auscultatory sound sensors 12 have not been processed, then the heart-cycle segmentation and heart-phase identification process 2300 is repeated beginning with step (2304) for the next auscultatory sound sensor 12. Otherwise, from step (2326), in step (2328), the heart-cycle segmentation and heart-phase identification process 2300 returns either the mean values ksi, ks2 of the corresponding root locations {ksi sTART, ksi END}, {ks2_START, ks2_END } associated with the corresponding SI and S2 heart sounds, or the corresponding mean values tsi, ts2 of the associated times, i.e. {t(ksi_sT \RT), t(ksi_END)}, {t(ks2_START), t(ks2_END)}, for each of the heart cycles 82 in each of the high-pass- filtered breath-held sampled auscultatory sound signals 72 from each of the associated auscultatory sound sensors 12.
Referring to FIG. 35, segmented high-pass-filtered breath-held sampled auscultatory sound data 72, s[] from each of the heart cycles 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9 may be synchronized with respect to the mean time ts2 associated with the S2 heart sound of each heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9, as illustrated in FIG. 35, or alternatively, may be synchronized with respect to the mean time tsi associated with the SI heart sound of each heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9, wherein the region of diastole extends from the time associated with the second root ks2_END , i.e. t(ks2_END), associated with the S2 heart sound, to the end of the associated heart cycle 82, 82.1, 82.2, 82.3, 82.4, 82.5, 82.6, 82.7, 82.8, 82.9.
Referring again to FIG. 21, either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72, is first analyzed within each region of diastole for each heart cycle 82 identified in step (2106/2300), and for auscultatory sound sensor 12, m to identify any outliers, and to detect, for selected pairs of auscultatory sound sensor 12, excessive noise. More particularly, beginning with step (2107), a heart-cycle pointer kuEAT is first initialized to a value of 1, after which, via step (2108), for each heart cycle 82 and for each auscultatory sound sensor 12, m, in step (2110), each region of diastole of either the breath-held auscultatory sound signal 16.1, or the corresponding high- pass-filtered breath-held sampled auscultatory sound signal 72, i.e. extending between samples is2[ksEG, /(BEAT] and isifksEG, keEAT+l] is checked for outliers, e.g. random spikes, using a beat outlier detection process 3600, wherein the heart-cycle pointer knFA r points to the particular heart cycle 82 within the selected breath-held segment ksEG, and is2[ksEG, kiiEA r] corresponds to the ks2_END of the associated heart cycle 82 at the beginning of diastole, and isifksEG, kBEAT+i ] corresponds to the end of diastole of that heart cycle 82, and the beginning of the systole for the next heart cycle 82. Referring to FIG. 36, the beat outlier detection process 3600 commences in step (3602) with receipt of the identifier m of the auscultatory sound sensor 12, the associated sampled auscultatory sound data Sm[], the segment counter ksEG, and the heart-cycle pointer 1¾EAT. Then, in steps (3604) and (3606), a standard-deviation counter ksTDDEv is initialized to a value of 1, and an index JMIN is set equal to is2|A.S£y„ ICBEAT] , the location of the beginning of diastole in the sampled auscultatory sound data Sm[], respectively. Then, steps (3608) through (3616) provides for repetitively calculating, in step (3612), a standard deviation STD[ksTDDEv] of the sample values Sm[jMiN : jMAx] in a plurality of windows of NWINDOW samples, each window shifted by one sample with respect to the next in steps (3614) and (3616)— that increment ICSTDDEV and jvnN, respectively,— resulting in a standard deviation array STD[] containing ksTDDEv standard deviation values, wherein, in step (3608), JMAX = jvnN + NWINDOW -1. For example, in one set of embodiments, NWINDOW is equal to 128. Then, in step (3618), the standard deviation array STD[] is rank ordered, and the median value thereof, MedianSTD, is used in step (3620) to calculate a standard deviation compactness metric STDDEVCM, as follows:
Figure imgf000044_0001
Then, in step (3622), if the standard deviation compactness metric STDDEVCM exceeds a threshold, for example, in one set of embodiments, equal to 6, but generally between 1 and 10, the particular region of diastole for the particular breath-held segment from the particular auscultatory sound sensor 12, m, is flagged as an outlier in step (3624). Then, or otherwise from step (3622), in step (3626), the process returns to step (2110) of the auscultatory sound signal preprocessing and screening process 2100.
Referring again to FIG. 21, then, in step (2112) of the auscultatory sound signal screening process 2100, if an outlier was detected in step (3622) and flagged in step (3624) of the beat outlier detection process 3600, then in step (2120), if the end of the breath-held segment has not been reached, i.e. if 1¾EAT < NBEATs(ksEG), then the auscultatory sound signal preprocessing and screening process 2100 repeats beginning with step (2108) for the next heart cycle 82.
Otherwise, from step (2112), referring again to FIGS. 17-20, the above-described noise detection process 1700 may be called from step (2114) to provide for determining whether or not a particular region of diastole for a particular heart cycle 82 of either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72, is corrupted by excessive noise, and if so, provides for flagging that region of diastole as being excessively noisy so as to be prospectively excluded from subsequent processing, wherein above references to breath-held auscultatory sound signal 16.1 in the description of the associated noise detection process 1700 should be interpreted as referring to the corresponding region of diastole, i.e. between samples is2|A.V£Y„ 1¾EAT] and isi |A.S£Y„ kBEAT+1] for the kBEAT th heart cycle 82 of segment ksEG of either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72, and references to the breath-held segment k in the associated noise detection process 1700 - which in this case would not be incremented in step (1726), with k\iAx=k in step (1724), - should be interpreted as referring to a corresponding particular heart cycle 82 associated with the particular region of diastole, wherein the noise detection process 1700 would terminate with step (1934) of the associated noise-content-evaluation process 1900.
Then, returning to the step (2116) of the auscultatory sound signal screening process 2100, if a noise threshold was exceeded in step (1930) of the noise-content-evaluation process 1900, in step (2120), if the end of the breath-held segment has not been reached, i.e. if keEAT < NBEATs(ksEG), then the process repeats beginning with step (2108) for the next heart cycle. Otherwise, from step (2116), the good beat counter GB is incremented in step (2118) before continuing with step (2120) and proceeding therefrom as described hereinabove. Otherwise, from step (2120) if the end of the breath-held segment has been reached, in step (2122), if a threshold number of valid heart cycles has not been recorded, the process repeats with step (2104) after incrementing the segment counter ksEG in step (2123). Otherwise, the recording process ends with step (2124).
Accordingly, each breath-holding interval B, ksEG of either the breath-held auscultatory sound signal 16.1, or the corresponding high-pass-filtered breath-held sampled auscultatory sound signal 72 is segmented into individual heart beats 82 (i.e. heart cycles 82, wherein reference to heart beats 82 is also intended as a short-hand reference to the associated breath-held sampled auscultatory sound data S[]) and the diastolic interval D is analyzed to determine the associated noise level to provide for quality control of the associated breath-held sampled auscultatory sound data S[], so as to provide for signal components thereof associated with coronary artery disease (CAD) to be detectable therefrom. Quality control of the recorded signals provides for detecting weak signals that may indicate health problems but can otherwise be blocked by strong noise or unwanted interference. The present method is developed for quantitative control of signal quality and can be deployed in the recording module 13 for real time quality monitoring or can be used at the post recording stage to extract only low-noise heart beats 82 that satisfy specific quality condition.
Referring to FIG.37, a second aspect of an auscultatory sound signal preprocessing and screening process 3700— also referred to as a beat selection algorithm, as was the first aspect auscultatory sound signal preprocessing and screening process 2100 - provides for collecting heart beats 82 that have diastolic noise level below the specific mean noise power level threshold Po. As described hereinabove, the associated auscultatory-sound-sensing process 700 proceeds through sequential acquisition of heart sound intervals with duration between 5 and 15 sec (for example, breath-holding intervals B). In step (3702), the sampled auscultatory sound data Sm[], the identifier m of the associated auscultatory sound sensor 12, and the associated segment counter ksEG are received, and, in step 3704), the corresponding recorded data block is passed through the heart-cycle segmentation and heart-phase identification process 2300 that identifies the beginning and the end of each heart beat 82 using synchronized ECG recording or another signal processing code that identifies timing of the first (SI) and second (S2) heart sounds, as described hereinabove. The second aspect auscultatory sound signal preprocessing and screening process 3700 input parameters also include the mean noise power level threshold Po and the required number NGMIN of high quality heart beats 82. In step (3706), a good beat counter GB is initialized to a value of 0. Following the heart-cycle segmentation and heart- phase identification process 2300, a two-dimensional array of heart beats 82 is created, for example, as illustrated in FIGS. 31 and 35. The heart-cycle segmentation and heart-phase identification process 2300 also identifies timing of diastolic interval D of each heart beat 82. Beginning with the selection of the first heart cycle 82 in step (3708), in step (3710), the breath- held sampled auscultatory sound data S[] is normalized with respect to absolute value of the S2 peak, and, in accordance with a first aspect, in step (3712), the mean noise power P is computed within a time window Tw. The time-window Tw is slid along the diastole D with the 50% overlap to compute an array of localized signal power P[] After the power P[iw] of the sampled auscultatory sound data Sm[] is calculated within each time-window Tw indexed by index iw, the maximum power PMAX of diastole D is determined i.e. as given by the maximum of the array of localized signal power P[], and, in step (3714), is compared against the mean noise power level threshold Po, the latter of which in one set of embodiments, for example, has a value of -20 dB. A noise power level P[iw] in one of the time-windows Tw greater than the mean noise power level threshold Po may indicate either excessive noise or presence of a large amplitude transient outlier.
Referring to FIG. 38, in accordance with a second aspect of heart-beat quality assessment procedure, the quality test consists of two stages designed to detect high amplitude transient spikes and noisy beats with broad-band spectrum. The quality tests are performed on the diastolic portion of the heart cycle 82, which is the region of primary interest for CAD detection. Nonstationary signals can be analyzed using a relatively short time sliding time-window Tw, for example, in one set of embodiments, with Tw = 30 ms (Nw = 120 samples at fs = 4 kHz), which is slid across diastole, with each time-window Tw advanced in time with respect to the previous time-window Tw by a stride period. In one set of embodiments, the stride period is equal to the duration of the time-window Tw, so that adjacent time-windows Tw abut one another without overlap. The total number of time-windows Tw (i.e. sub-segments) in diastole, without overlap, is given by Ns = N/Nw, where N is the total number of samples during diastolic, and Nw is the number of samples in each time-window Tw.
The variance within each time-window Tw - used to detect any outliers therein - is computed as:
Figure imgf000047_0001
wherein Xu and mc are respectively the kJh sample and the mean value, of the i,h time-window Tw, respectively. The local signal power of the th time-window Tw, is given by:
T, = s (16)
An outlier power threshold PLIM is determined by adding 6 dB to the median value of Pi for all time-windows Tw, and in step (3714), if the value of Pi exceeds PLIM for any time-window Tw, then, in step (3720), the current heart cycle 82 is ignored.
If none of the time-windows Tw include an outlier, then the mean power of diastole is given by:
Figure imgf000047_0002
The associated noise power threshold Rtΐ is defined with respect to the 2-byte A/D converter range, so that:
Th
P 10
1 Th = 1 P ADC -110 u (18) wherein Th is a predetermined threshold, for example, -50 dB, and PADC is the power associated with the maiximum signed range of a 2-byte A/D converter, i.e. (32767)2. Accordingly, if, in step (3714), the mean diastole power Pm exceeds the noise power threshold Pn, then, in step (3720), the current heart cycle 82 is ignored.
If, from step (3714), the diastolic signal power exceeds the mean noise power level threshold Po, then, in step (3720), the associated heart beat 82 is labeled as noisy beat and is not counted in the overall tally of heart beats 82. Otherwise, from step (3714), the good beat counter GB in step (3716), and if, in step (3718), if the number of good heart beats 82, i.e. the value of the good beat counter GB, is less than the required number NGMIN of high quality heart beats 82, then the second aspect auscultatory sound signal preprocessing and screening process 3700 repeats with step (3708) for the next heart cycle 82. In one set of embodiments, if the required number of high quality heart beats 82 is not reached within a reasonable period of time, then the user is informed that recording is excessively noisy so that additional actions can be performed to improve signal quality.
Following the acquisition of a sufficient number of sufficiently-low-noise heart beats 82, following step (3718), the sampled auscultatory sound data Sm[] is further preprocessed to emphasize acoustic signals in diastole D and extract signal specific features that can be used for disease classification. For example, in step (3722), the mean position of the peak of the S2 heart sound - for example, the above-described mean value ts2 -- is determined for each heart beats 82, after which the heart beats 82 are synchronized with respect thereto, for example, as illustrated in FIGS. 39a and 39b, which respectively illustrate a stack of heart beats 82 before and after this S2 alignment, wherein the synchronization of the heart beats 82 with respect to the S2 heart sound helps to locate an acoustic signature that might be present in a majority of the heart beats 82 and that is coherent between or amongst different heart beats 82.
Then, in step (3724), the heart beats 82 are normalized with respect to time so as to compensate for a variation in heart-beat rate amongst the heart beats 82, and to compensate for a resulting associated variation in the temporal length of the associated diastolic intervals. Generally, heart-beat rate is always changing and typically never remains the same over a recording period of several minutes, which if not compensated, can interfere with the identification of specific signal features in diastole D, for example, when using the below-described cross-correlation method. Although heart-beat segmentation alone provides for aligning heart-beat starting points, variations in the heart-beat rate can cause remaining features of the heart cycle 82 to become shifted and out of sync - i.e. offset— with respect to each other. However, such offsets can be removed if the associated heat beats 82 are first transformed to common normalized time scale t/T*, where T* is the fixed time interval, for example, the duration of the slowest heart beat 82, followed by beat resampling and interpolation so as to provide for normalizing the original signal at a new sampling rate, for example, as illustrated in FIGS. 40a and 40b that respectively illustrate stacks of heart beats 82 before and after such a time normalization, wherein FIG. 40a illustrates complete heart cycles 82, and FIG. 40b illustrates only the temporally -normalized diastolic portions D thereof.
Referring again to FIG. 37, in step (3730), a final stage of signal pre-processing provides for extracting acoustic features from the recorded heart beats 82. There are numerous options that can be applied to feature extraction which typically involves certain transformation of raw data to low-dimensional or sparse representation that uniquely characterizes the given recording. Alternatively, the raw data can be transformed in a set of images and some image recognition algorithm like convolutional neural net can be employed for automatic feature selection. For example, in accordance with one set of embodiments, referring to FIG. 41, a local cross- correlation (CC) of multiple heart beats 82 provides for identifying a relatively high-pitch signal component in the diastolic phase of the heart cycle 82 occurring in multiple heart beats 82, which can be result from turbulent blood flow, wherein pairs of heart beats 82 of a 2-D stack of heart beats 82 - each segmented from the above-described high-pass-filtered breath-held sampled auscultatory sound data 72, s[]— are cross-correlated with one another by computing an associated set of cross-correlation functions Rxixj for each pair of heart beats 82, xi[], xj[]. This computation is made using a sliding short-time window with Nw samples (for example, typically 128) which is advanced in time one sample per each iteration of the cross-correlation computation. The cross-correlation is computed without time lag, resulting in an array - of the same size as that of the original signals— that is given by:
Figure imgf000049_0001
The cross-correlation assigned to each beat is given by an average of cross-correslations thereof with the remaining Nb - 1 heat beats 82:
Figure imgf000049_0002
wherein xi and xj are the diastolic high-pass-filtered breath-held sampled auscultatory sound data 72, s[] of two distinct heart beats 82, and Nb is the total number of heart beats 82 in the 2- D stack. Following computation of all possible pairs of heart beats 82, an NbxNt cross-correlation matrix is obtained and displayed as a 2D image, wherein Nt is the number of time samples in each heart beat 82. A similar signal pattern during diastole that is present in majority of heart beats 82 will produce a localized cross-correlation peak in diastole. Accordingly, cross-correlation peaks associated with a micro-bruit signal occurring at approximately at the same temporal location within the diastole interval from one heart beat 82 to another will produce distinct bands across the image within the same temporal region of each heart beat 82, for example, as illustrated in FIG. 42, which illustrates an image of the cross-correlation function Rxi as a function of n (corresponding to time) for each heart beat 82, with different heart beats 82 at different ordinate positions, wherein the value of the cross-correlation function is indicated by the color of the associated pixels. The cross-correlation operation provides for emphasizing signal features that are coherent within the current time window, and provides for suppressing uncorrelated noise, thereby providing for increasing the associated signal-to-noise ratio.
Alternatively, or additionally, acoustic features in diastolic interval of heart beat 82 can be visualized using continuous wavelet transform (CWT). The wavelet transform processing is similar to the short-time cross-correlation but instead of cross-correlating signals from different heart beats 82, the signal of interest is correlated using a wavelet function with limited support to facilitate temporal selectivity.
Figure imgf000050_0001
wherein the associated mother wavelet y is a continuous function in time and frequency domains. The wavelet family is typically chosen empirically using specifics of the signal under consideration. The family of Morlet wavelets appears as an appropriate choice for analysis of heart sounds. Output of the wavelet transform is a two-dimensional time-frequency representation of signal power |X(a,b)|2 defined in terms of scaling and shift parameters. An example of a wavelet transform - using a 6th order Morlet wavelet— is illustrated in FIG. 43, where the associated corresponding original color map represents distribution of the signal power in a time-frequency plane, with normalized time-shift - as a function of b— as the abscissa, and frequency as the ordinate, with frequency f in Hertz given as follows:
Figure imgf000050_0002
wherein w is the wavelet order, for example, w=6 for a 6th order Morlet wavelet, and the results plotted in FIG. 43 are the result of a cumulative processing of the diastolic portions of a plurality of heart cycles. For the purpose of signal classification, the wavelet representation can be reduced further by dividing the time axis into discrete intervals and computing overall signal power within such interval and specific frequency bandwidth. For example, in one embodiment, the wavelet image may be subdivided in time intervals of 200 milliseconds and two frequency bands of 10 Hz - 40 Hz and 40 Hz - 100 Hz. The resulting output vector of the signal power within the associated intervals of time and frequency can be used as an input to a neural network classifier.
Generally, step (3730) of the second aspect auscultatory sound signal preprocessing and screening process 3700 - which may also be used in cooperation with the above-described first aspect auscultatory sound signal preprocessing and screening process 2100 - incorporates one or more feature extraction algorithms which identify significant signal parameters that can be linked to coronary artery disease (CAD) and which can be used for training a machine learning algorithm for automatic CAD detection. Furthermore, if auscultatory sound signals 16 are recorded at a relatively high sampling rate (for example, at a 4 kHz sampling rate), each heart beat 82 might contain over 4000 samples per each of six channels. Such large amount of highly correlated variables makes the usage of the raw waveform for classification very difficult without additional signal processing to reduce the dimensionality of the problem. Such dimensionality reduction can be achieved by use of an appropriate feature extraction algorithm that identifies a reduced set of parameters that are related to CAD. Mathematically, the feature extraction procedure provides a mapping of the high-dimensional raw data into the low-dimensional feature space with adequate inter-class separation. For example, standard dimensionality reduction routines such as singular value decomposition (SVD) or principal component analysis (PCA) may be used to decompose raw data onto orthonormal basis and to provide for selecting relevant features with minimal loss of information. The time domain signal itself can be transformed prior to feature extraction to emphasize unique features thereof. For example, frequency domain representation by Fourier transform can be advantageous for feature extraction if the signal contains discrete set of characteristic frequencies. The performance of a signal classifier can be dramatically improved by excluding a large number of irrelevant features from analysis. In accordance with one aspect, the signal classification problem begins with a mapping from the original high-dimensional space (size N) to a feature space (size p « N), followed by a mapping of the feature space to an m-dimensional space, wherein the dimension m is equal to the number of classes. For example, for a binary classification problem— e.g. CAD or no CAD,— m=2.
In accordance with one set of embodiments, step (3730) of the second aspect auscultatory sound signal preprocessing and screening process 3700 employs a wavelet packet transformation (WPT) for sparse representation of heart sounds in time-frequency domain, followed by a custom designed binary classifier. Several standard classifier algorithms can be trained using reduced feature set, and to provide for binary classification of the associated heart sounds- useable either individually or in combination,— including , but not limited to, a support vector machine (SVM), a fully -connected artificial neural network (ANN), or a convolution neural network (CNN) applied to two-dimensional time-frequency images.
Referring to FIGS. 57-66, a wavelet packet transformation (WPT) processing stage provides for reducing dimensionality by converting raw time-domain auscultatory sound signals 16 into a time-frequency basis using discrete wavelet transform (DWT), followed by elimination of associated components that do not provide significant contribution to the original signal or do not provide substantial contrast between two classes of interest. For example, referring to FIG. 57, in one set of embodiments of discrete wavelet decomposition, an initial (input) signal x[n] is passed through a series of stages at which low-pass and high-pass filter functions (quadrature mirror filters) are applied to obtain approximation iij(k) and detail dj(k) coefficients at the /* level of decomposition. This procedure is repeated recursively at the subsequent steps with the only approximation coefficients used as an input, The coefficients generated at each indicated level of filtering provide the corresponding amplitude of the initial signal x[n] after filtering by a filter having a corresponding bandpass filter response characteristic illustrated in FIG. 58.
Referring to FIGS. 59 and 60, In accordance with one set of embodiments, the discrete wavelet transform (DWT) is implemented using a Daubechies wavelet family, for example, a Daubechies 4 (db4) wavelet family, for which the associated scaling function f— having low- pass filter coefficients ho through h — is illustrated in FIG. 59, and for which the associated wavelet function Y— having high-pass filter coefficients go through - is illustrated in FIG. 60.
Referring to FIGS. 57 and 61, at each node 92 of the discrete wavelet transform (DWT), an input time series Wj.kfl/ from decomposition level j is transformed into two output time series at decomposition level /+/. i.e. Wj+i,2k[l] and Wj+i,2k+i/l/. each containing half the number of samples as in the input time series, and mutually-exclusive halves of the frequency content of the input time series, wherein the transformation to generate the Wj+i^k+ifl] lower-half bin of frequency content is given by the transformation 61.1g 61.2g illustrated in FIG. 61 and in equation 22 below, and the transformation to generate the Wj+i,2k[l] upper-half bin of frequency content is given by the transformation 61.1h 61.2h illustrated in FIG. 61 and equation 23 below, wherein k designates an associated frequency bin for the associated decomposition level j.
The filter functions are designed to provide for energy conservation and lossless reconstruction of the original signal from the set of transformed time series Wj.kfl/ from a particular decomposition level j. These properties along with smoothness requirements define the family of scaling and wavelet functions used for decomposition. The resulting set of K=2' distinct frequency bands at each decomposition level j, together with the corresponding associated transformed time series W/ /// can be used for analysis and feature extraction, instead of relying upon the corresponding original raw signal xfnj.
The wavelet packet transformation (WPT) is generalization of the standard multi-level DWT decomposition, wherein both approximation and detail coefficients are decomposed using quadrature mirror filters, for example, as described in M. Wickerhauser,“Lectures on Wavelet Packet Algorithms”, http://citeseerx.ist.psu.edu, which is incorporated herein by reference. FIG. 62 illustrates a wavelet packet transformation (WPT) at decomposition level j=3, and FIG. 63 illustrates the bandpass frequency responses associated with each of the outputs thereof. Analytically, the wavelet packet transformation (WPT) decomposition can be described by the following recursive expressions:
Figure imgf000052_0001
where gm is the coefficient of the scaling function and hm is the coefficient of the wavelet function, k is the index of the associated frequency bin at decomposition level j, and l is the index of the associated time series array associated with the particular frequency bin k. The wavelet packet transformation (WPT) provides a benefit of sparse signal representation similar to the discrete wavelet transformation (DWT), but also provides better resolution of frequency components by decomposing the detail part of discrete wavelet transformation (DWT), which results in the sub-band structure illustrated in FIG. 63. The maximum decomposition level j for a given signal depends on the signal length and the trade-off between desired time and frequency resolutions. The resulting tree-like structure of wavelet packet transformation (WPT) bases - each base corresponding to a particular frequency bin A at a particular decomposition level j— provides a highly redundant representation of the original signal x[n], which limits the amount of dimensionality reduction. However, heart sounds have characteristic properties that can be well described by a relatively small number of wavelet packet transformation (WPT) bases and a substantial amount of irrelevant information can be discarded. For example, in one set of embodiments, the WPT tree nodes with highest energy concentration are retained for further analysis and classification. For example, FIG. 65 illustrates a wavelet packet transformation (WPT) energy map of the high-pass-filtered breath-held sampled auscultatory sound signal 72 for the heart cycle 82 illustrated in FIG. 64, generated by an 8- level wavelet packet transformation (WPT) using a Daubechies 4 (db4) wavelet family, wherein the individual nodes 92 - or associate bases or frequency bins— are outlined with rectangular boundaries in the image, and each frequency bin is color coded to indicated the total energy Ejk of the kJh node at the/* decomposition level, which is computed using wavelet packet transformation (WPT) coefficients WjkflJ as follows:
Figure imgf000053_0001
For each decomposition level j, the total energy from all frequency bins is the same, i.e.
ETotai
Figure imgf000053_0002
s the same value regardless of decomposition level j. (25) k
The wavelet packet transformation (WPT) energy map of FIG. 65 illustrates that most of the signal energy is concentrated in the relatively low frequency bands, and that only the levels 7 and 8 of decomposition show the fine structure of the energy distribution. Accordingly, this example shows that most of the wavelet packet transformation (WPT) nodes 92 can be ignored without significant loss of information. A formal approach to the selection of the best bases— that relies on Shannon entropy as a cost function computed for each node in the tree—is described by H. Saito and R. Coifman in the following documents that are incorporated herein by reference: N. Saito, R.R. Coifman, “On local feature extraction for signal classification”, https://www.math.ucdavis.edu/ ~saito/publications/saito_iciam95.pdf, and N. Saito, R.R. Coifman,“Local discriminant bases and their applications”, J. Math. Imaging and Vision, 5, 337- 358 (1995). For example, traversing the wavelet packet transformation (WPT) tree from the leaf level up (i.e. highest to lowest decomposition level j), children nodes with the combined entropy lower than that of the parent node will be retained as the most relevant for a given signal, for example, as illustrated in FIG. 66 for the wavelet packet transformation (WPT) energy map illustrated in FIG. 65, wherein the Shannon entropy is given by:
Figure imgf000054_0001
The Wavelet Packet Transform (WPT) provides a redundant representation of actual signal using the discrete wavelet transform with various decomposition levels that provide higher frequency resolution (in terms of filter banks) at the expense of reduced time resolution. The WPT of the signal can be described by a set of coefficient wj(k, 1), where j is the decomposition level, k is the filter bank index and 1 is the coefficient index within filter bank. An example of WPT decomposition is shown in FIG. 65 where each cell outlines positions of filter bank for each level and intensity is related to the signal energy from equation 24. It can be seen that energy distribution wi thing specific frequency bands becomes apparent at high decomposition levels ( j > 6). Although, the color of the high frequency cells is dark blue, the values of the associated WPT coefficients is not zero.
The entire WPT energy map is highly redundant and not well suited in its raw form for signal classification. Instead, only those cell that uniquely characterize a given signal are kept while ignoring remaining cells. This can be accomplished using the best WPT basis selection algorithm which employs entropy WPT coefficients as a metric of the filter bank (cell) importance. This algorithm runs from leave cells up along the decomposition tree and selects parent cell whose entropy is less compared to the sum of the two children cells. An example of best basis selection applied to the WPT map in FIG. 65 is shown in FIG. 66. Values the best basis would generally not coincide with the energy map shown in FIG. 65 because entropy and energy differ from one another,
In FIG. 65, the total energy of a particular cell is represented by the intensity thereof. FIG. 65 illustrates that most of the energy is concentrated in the relatively -lower frequency bands. The cells of FIG. 66 - for which intensity is inversely related to entropy - are considered to be a signature of a given signal, and do not overlap with the corresponding cells of FIG. 65, the latter of which is representative of energy. None of the cells below the highlighted (children) cells in FIG. 66 are selected because they exhibit high entropy.
The wavelet packet transformation (WPT) energy map and the associated best basis selection can be used to reduce dimensionality of the heart beat classification problem by analyzing the signal represented by transformed time series W/A/// and rejecting information irrelevant for the classification task. The wavelet packet transformation (WPT) is one of a variety of signal processing techniques that can be used for extraction of important parameters or features suitable for prediction of CAD. The very basic set of features may include typical metrics of raw signals (amplitude, timing, spectral power and other) that can be derived from segmented heart beats. However, such hand-crafted feature set may not be optimal for current problem of CAD classification. Regardless of which method is used for feature extraction, the output of this data processing stage is a vector of p elements with p « N, where N is the size of raw signals. The feature vector can be represented either as a 1-D array of p elements or as a 2-D matrix for a classification algorithm operating on image information. There are several powerful classification algorithms that can be trained for disease detection using recorded heart sounds and the extracted feature vector. These algorithms include support vector machine (SVM), feed-forward artificial neural network (ANN) and convolutional neural network (CNN), which is particularly suitable for 2-D image classification.
Referring to FIG. 67, a support vector machine (SVM) is a powerful supervised machine leaning algorithm frequently used for data classification problems, which has several useful properties and can operate reliably on a small data sets with poor class separation. In case of linearly separable data, the SVM classifier will create decision hyperplane in feature space with highest separation between two classes using training data set. For example, FIG. 67 illustrates a trained SVM classifier and decision boundary for a two-dimensional feature space with two classes. The support vectors are circled, and the highest margin hyperplane is shown by the solid straight line. When separation between classes is less then perfect, the tolerance of SVM algorithm to data misclassification can be tuned by adjusting an associated C parameter during training. In more complicated situations, a nonlinear decision boundary can be created using a kernel function to project data into higher dimensions and apply SVM algorithm to modified data.
The SVM algorithm can be used as a CAD classifier with data recorded by the Data Recording Application (DRA) 14, 14.1. Prior to sending data to SVM algorithm, recordings are processed by beat segmentation and feature extraction stages, to produce a feature vector for each test-subject 22, by preprocessing of n available recordings and extracting p features, with each channel data then transformed into an n x p feature matrix. The feature vector extracted from the segmented beats can be either a set of custom selected metrics (amplitude, timing of specific segment, energy, statistic parameters, sample entropy and others) or a subset of wavelet packet transformation (WPT) coefficients associated with the signal region of interest. If the resulting number of features p is still relatively high, a principal component analysis (PCA) procedure can be applied to identify a subset of features with highest variance and eliminate correlating features. After all preprocessing steps and data normalization, the feature matrix is split into testing and training sets with ratio 1 to 4 respectively. The training set is used to train SVM and optimize classifier hyper-parameters, while the testing set is used to evaluate classifier performance with unseen data. Computer code that provides for implementing a SVM classifier is available in several open source packages for Python and R programming languages, for example, the skleam machine learning package in Python.
Referring to FIG. 68, a feed-forward artificial neural network (ANN) provides an alternative option for classification of auscultatory sound signals 16 following preprocessing and feature extraction stages. An artificial neuron is a nonlinear processing element with p inputs and a nonlinear activation function that generates an activation output. Assuming x(p) is a feature vector which is sent to the p inputs of each neuron, then the associated activation output of ilh neuron is computed as:
Figure imgf000056_0001
wherein wy is the weight matrix that defines connection strength of i,h neuron to fh input, /;, is the neuron bias and g(z = wTx + b) is the associated nonlinear activation function. The specific form of the activation function g is chosen at the design stage, wherein commonly used functions include sigmoid, hyperbolic tan and rectified linear unit (ReLu). The network of interconnected neurons constitutes the artificial neural network (ANN), which can be capable of modeling relatively complicated relationships between the input vector and target class variables by adjusting weights and biases during training stage.
Properties of a specific artificial neural network (ANN) implementation are defined at the design stage and include: 1) number of hidden layers, 2) number of neurons per hidden layer, 2) type of activation function, 3) learning rate and 4) regularization method to prevent overfitting. For example, in one set of embodiments, the artificial neural network (ANN) is implemented using the open source TensorFlow deep learning framework, which provides for setting each of these parameters. The neuron connection strength is defined by the weight matrix wy for each layer which is adjusted during network training and a cost function evaluated at each training epoch using available truth labels. The artificial neural network (ANN) training is accomplished by a standard back-propagation algorithm using a cross-entropy cost function. The network design specifics are determined by the available data since a network with multiple hidden layers (deep ANN) can be very powerful but also is prone to overfitting when a small data set is used. Therefore, the specific artificial neural network (ANN) architecture is developed on a trail-and- error basis, dependent upon the available data and the size of feature vector. In one set of embodiments, output layer of the artificial neural network (ANN) for a binary classifier is implemented as the softmax function with two-element vector [1, 0] for CAD positive and [0, 1] for CAD negative case.
For example, in one set of embodiments, the features used as inputs to either an ANN classifier or a SVM classifier are selected from the following: length (time duration) of: S 1 , systole, S2, and diastole; ECG RR peak duration; amplitudes of SI, systole, S2, and diastole; ratios of corresponding amplitudes; and energy of spectral bands associated with SI, systole, S2 and diastole. Other features such statistics (mean, variance, skewness and kurtosis) of the parameter distributions can also be used as features for input to either an ANN classifier or a SVM classifier.
Referring to FIG. 44, in one set of embodiments, a convolutional neural network is employed to analyze cross-correlation images of acoustic signals in order to provide for automating the process of acoustic feature extraction. The network input consists of the three- dimensional array of cross-correlation data recorded from the left or right positions on the thorax 20 of the test-subject 22, and combined into a single array, for example, representing three individual channels associated with three different auscultatory sound sensors 12. The convolutional neural network (CNN) comprises several convolutional layers that each use a 5x5 kernel array to scan input data to build a structure of acoustic features with increasing complexity. Neuron activation uses rectified linear unit (ReLU) nonlinearity to produce output data. The final stage of the network classifier is a fully connected neural net with an additional clinical information (age, gender, blood pressure, etc.) merged with the acoustic features identified by CNN. The network output is a two-node layer implementing binary classification via softmax function applied to the incoming data. The CNN classifier is trained using 80% of all available data with binary CAD labels (Yes = 1 and No = 0). The remaining 20% of data is randomly selected to test the classifier performance. Reported performance metrics include prediction accuracy, sensitivity, specificity, negative prediction value and positive prediction value. New patient data are classified by passing through the same pre-processing stages and feeding the computed cross-correlation image to CNN classifier.
Convolutional neural networks (CNN) have proved to be very efficient for prediction and classification problems especially with large scale problems involving images. The typical size of input image data can be quite large, which makes application of standard feed forward networks either impractical or even impossible due to huge number of parameters to be trained. Convolutional neural networks (CNN) accommodate the size of the problem by weight sharing within small number of neurons comprising a receptive field that is scanned over the 2-D input. One benefit of using a convolutional neural network (CNN) for machine learning problems to the ability thereof to leam important features directly from data, so as to provide for bypassing the feature extraction stage that is used by support vector machine (SVM) and feed-forward artificial neural network (ANN) classifiers. A typical convolutional neural network (CNN structure consists of one or more convolution and pooling layers that build a hierarchical structure of features with increasing complexity. Following convolution and max pooling, the extracted features are fed to a fully connected network at the final stage of convolutional neural network (CNN) classifier. For example, FIG. 44 illustrates a convolutional neural network (CNN) architecture incorporating a single convolution layer. Each cell of the max pool contains the maximum value of a corresponding array of cells in the corresponding convolution layer.
The receptive field is a relatively small 2-D array of neurons (for example, 5 x 5) that is scanned across the input image while performing an associated cross-correlation operation. A relatively small number of connected neurons provides for a relatively small number of corresponding weights to be adjusted. The max polling operation provides for reducing the size of the input to the associated fully-connected neural network by selecting pixels with maximum intensity from the associated convolution layer. Similar convolution and max polling operations can be performed multiple times to extract the most significant features before submitting to an associated fully -connected neural network for classification. Although a convolutional neural network (CNN) can be trained to recognize complicated patterns in 2-D data sets, this typically requires large amount of data for efficient generalization and to avoid overfitting. The convolutional neural network (CNN) classifier can be applied either directly to the auscultatory sound signals 16 (or filtered versions thereof), or to corresponding 2-D images generated therefrom, for example, using either a continuous wavelet transform or an associated decomposition thereof by wavelet packet transformation (WPT) . For example, in cooperation with the extraction of features using wavelet packet transformation ( WPT), the coefficients of Jlh level of decomposition can be transformed into a matrix with dimensions (N/2J) x 2J. where N is the size of the time domain signal. Such 2-D data can be used train the convolutional neural network (CNN) classifier in order to find any patterns associated with CAD. This type of network is more complicated than a standard feed-forward artificial neural network (ANN), and utilizes more hyperparameters to be tuned to achieve optimal performance. In addition to parameters applicable to an artificial neural network (ANN), the convolutional neural network (CNN) design includes specification of the number of convolution layers, the size of receptive fields (kernel size), number of channels processed simultaneously, filter properties, regularization. After finalization of its design, the convolutional neural network (CNN) can be trained using a training data set and then evaluated using an unseen test data set. For example, for one set of embodiments, the open- source TensorFlow flexible deep learning toolkit and API - which provide for building high- performance neural networks for variety of applications - have been used to design and train the convolutional neural network (CNN) for detecting CAD. Referring to FIG. 72, a second embodiment of a convolutional neural network (CNN) incorporates a plurality of convolution stages (Conv-1, Conv-2, Conv-3, Conv-4 and Conv-5) that are interspersed with corresponding associated MaxPool stages, wherein the output of the final max pooling stage is fed into a first fully -connected neural network (FC1) with 128 neurons, the output of which is fed into a second fully-connected neural network (FC2) with 64 neurons, the output of which is processed by a SoftMax output operation given by:
Figure imgf000059_0001
where zi = wyxj + bi , bi is a bias term, and repeated indexes imply summation. The result of the classification (CAD/No CAD) is defined by the highest value of output ou .
The input to the convolutional neural network (CNN) is a series of either FIGS, 41/42-sytle convolution images, or a series of FIG. 43-sytle CWT transform images, with one image for each of the six different auscultatory sound sensor 12, with each image, for example, given by a 100 by 200 array of pixels.
Each of the convolution stages (Conv-1, Conv-2, Conv-3, Conv-4 and Conv-5) provides for convolving the array of input elements (or pixels) with an either 3x3 or 5x5 kernel or weight array, which is scanned across the input image with a stride if one pixel. If zero padding is used,, a single kernel generates one feature map with the size as the original image, but 2D convolution applied as follows:
Figure imgf000059_0002
where x(kl, k2) is the input image, and w is the kernel matrix. To extract a variety of features, each convolution layer has multiple kernels that are applied to the input image. For example, in FIG. 72, each convolution stages (Conv-1, Conv-2, Conv-3, Conv-4 and Conv-5) is designated as P x MxN K=LxL, wherein P is the number of different kernels that are applied, MxN is the dimension of the associated image, and LxL is the dimension of the associated kernel matrix. If zero padding is not used, the dimension of the output from the convolution stages (Conv-1, Conv- 2, Conv-3, Conv-4 and Conv-5) is reduced by (L-l). Following convolution, the max pooling layer is applied which is a simple mask 2x2 scanned across feature maps. At each position only one pixel with highest magnitude is selected which is essentially a decimation operation that reduces the size of feature map by 2 while selecting brightest pixels.
It should be understood that the CNN, ANN and SVM classifiers each provide an independent means of classifying the associated input features in order to determine whether or not CAD is likely. These are alternative classification algorithms and a decision as to which one to use is typically based on their respective performance, for example, prediction accuracy. Alternatively, the different algorithms could be combined to make a single prediction, for example, using either a voting scheme or a weighted sum of likelihoods.
The auscultatory coronary-artery-disease detection system 10 can present various views of the acoustic data (both unprocessed and processed) that was captured during the test for review by a clinician. By reviewing different visualizations of the test results, the clinician may strengthen his case for a particular diagnosis, either in agreement with or disagreement with the result produced by the system. Additionally, particular visualizations may help reassure the patient that the diagnosis is the correct one.
Referring to FIG. 45, the system presents heart/arteries view comprising a graphical representation of the heart and coronary artery tree, highlighting the data in accordance with a Textual View. By selecting any of the occlusions, the clinician is presented with the option to highlight the blockage in either: a Stacked Heartbeat View; a Bruit Identification View (either mode); a Bruit Analysis View; or a Receiver Operating Characteristic Curve. For example, FIG. 45 illustrates test results showing severity of obstructions and zonal location of each obstruction within coronary artery tree. The tree diagram also indicates which arteries are critical, and which are not, with associated color coding in accordance with the amount of occlusion, for example, with the illustrated region of 50 occlusion highlighted in a reddish hue, and the region of 30% occlusion highlighted in a yellowish hue.
In accordance with the Textual View, the system presents, in a textual manner, the system’s interpretation of the acoustic data, including: whether or not the patient has a clinical level of CAD; the count of obstructions detected, for each obstruction: the location (zone) of the obstruction; the percentage occlusion of the obstruction; or the type of blockage (soft plaque, hard plaque). For each of the items in the view listed above, the system provides for presenting an associated confidence level, which indicates how confident the system is of each specific element of information presented, or, based on the patient’s demographic data (age, sex, BMI, medical and family history), the percentage of other patients who have a similar count, severity, and position of obstructions. From this view, the clinician may switch to any other view listed in this document for more information.
Referring to FIG. 46a, the system presents a Receiver Operating Characteristic (ROC) curve to highlight the system’s CAD detection algorithm sensitivity as a function of its false positive rate (1 -specificity). The graph’s attributes will be as is standard for ROC curves - vertical axis representing sensitivity, horizontal axis representing the false positive rate, with a 45-degree dotted line representing a“coin flip” test (AUC = 0.5).
On this graph, the system will plot the ROC curve and calculate the Area Under the Curve (AUC) based on the patient’s demographics, and the system’s clinical trial results. The coordinate that corresponds to the current positivity criterion will be highlighted. The clinician will be able to display a modified graph if he commands the system to exclude specific parts of the patient’s demographic data. The graph may or may not contain gridlines and/or data points defining the ROC curve, and it may or may not fill in the AUC.
Alternatively, referring to FIG. 46b, the data may be visualized as a line graph (with or without underfill), where true positives, true negatives, false negatives, and false positives are displayed, with a line indicating the currently-used cutoff value.
Referring to FIGS. 47 through 53, the system provides for visualizing the acoustic data resulting from an analysis of the breath-held auscultatory sound signals 16.1 from the auscultatory sound sensors 12.
Referring to FIG. 47, the system presents a heartbeat view comprising a graphical plot of the systolic and diastolic intervals of each heartbeat captured. The horizontal axis of the graph represents captured heartbeats. This representation is a subset of the acoustic capture time, as some acoustic data acquired is discarded (poor quality, etc.) during the system’s analysis process. The duration of the SI and S2 sounds are also highlighted on this axis.
Generally the vertical axis may comprise acoustic data captured from each of the system’s sensors - both acoustic and ECG. A correlation procedure is performed to ensure that the data captured from each of the system’s sensors is aligned to one another. Because the initial display of this view may contain many dozens of heartbeats, the clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, and explore any discrepancies between the data captured by the ECG and acoustic sensors for any particular heartbeat. For example, FIG. 47 illustrates an example of the Heartbeat View wherein the acoustic view has been limited by the clinician to Sensor#3 only, and zoomed in to see only 2 heartbeats, with the ECG data also displayed.
Referring to FIG. 48, in accordance with a Stacked Heartbeat View, the system presents a graphical plot of the systolic and diastolic intervals of each heartbeat captured. The horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the systolic interval to the end of the diastolic interval. The duration of the SI and S2 sounds are highlighted on this axis. The vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture, where each heartbeat is itself a graph, with a vertical axis representing intensity. A correlation procedure is performed to ensure that the start of the systolic interval for each heartbeat is aligned on x-axis=0. For each heartbeat, the system highlights any unexpected acoustic signals captured, as such signals may be an indication of an obstruction or other cardiac condition. For example, in FIG. 48, unexpected acoustic signals are highlighted in red. The clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
Referring to FIG. 49, in accordance with a Bruit Identification View in a Line Graphs with Underfill mode, the system presents a graphical plot of unexpected acoustic signals captured during the diastolic interval by means of a line graph with underfill. The horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the diastolic interval to the end of the diastolic interval. The duration of the S2 sound is also highlighted on this axis. The vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture, where each heartbeat is itself a graph, with a vertical axis representing intensity. A correlation procedure is performed to ensure that the start of the diastolic interval for each heartbeat is aligned on x-axis=0. An underfill is used to visually highlight deviation from the baseline.
For each heartbeat diastole displayed, the system highlights any unexpected acoustic signals captured, as such signals may be an indication of an obstruction or other cardiac condition. Such signals are highlighted by the height of the line graph, which represents intensity. These highlighted areas on the graph allow a clinician to distinguish low-energy noise (which may or may not be a sign of non-obstructive CAD) from high-energy noise (which is a likely indicator of obstructive CAD). In this view, it is also easy for a clinician to understand the correlation (across heartbeats) of noise, as well as the number and timing of correlated instances of noise (which may indicate the number and location of blockages, respectively). For example, the data illustrated in FIG. 49 exhibits relatively high intensity noise in many heartbeats at approximately t=0.05 seconds into diastole.
The clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
Referring to FIG. 50, in accordance with a Bruit Identification View in a Spectrogram mode, the system presents a graphical plot of unexpected acoustic signals captured during the diastolic interval by means of a spectrogram. The horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the diastolic interval to the end of the diastolic interval. The duration of the S2 sound is also highlighted on this axis. The vertical axis is comprised of the acoustic data captured for each of the heartbeats, in either ascending or descending order of capture. A correlation procedure is performed to ensure that the start of the diastolic interval for each heartbeat is aligned on x-axis=0.
For each diastole displayed, the system highlights any unexpected acoustic signals captured, as such signals may be an indication of an obstruction or other cardiac condition. Such highlighted areas indicate the intensity of the unexpected signal. Highlights are in the form of color, which indicate varying intensity of the signal. This view could alternatively be represented in monochrome as a contour plot. For example, the data illustrated in FIG. 50 exhibits relatively high intensity noise in many heartbeats at approximately t=0.05 seconds into diastole.
These highlighted areas on the graph allow a clinician to distinguish low-energy noise (which may or may not be a sign of non-obstructive coronary artery disease (CAD() from high- energy noise (which is a likely indicator of obstructive CAD). In this view, it is also easy for a clinician to understand the correlation (across heartbeats) of noise, as well as the number and timing of correlated instances of noise (which may indicate the number and location of blockages, respectively).
The clinician has the option to highlight a subset of the data on the horizontal axis, and command the system to zoom into the selected section. In this way, the clinician can perform a deeper analysis on one or more sections of the acoustic capture data as he so chooses, especially so that he may explore more deeply any unexpected acoustic signals.
Referring to FIG. 51, in accordance with a Bruit Analysis View, the system presents a graphical time/frequency plot of unexpected acoustic signals captured. The horizontal axis of the graph represents time (in seconds or milliseconds) from the beginning of the systolic interval to the end of the diastolic interval. The duration of the SI and S2 sounds are highlighted on this axis. The vertical axis represents the frequency (in Hz or kHz) of any unexpected acoustic signals, averaged from all captured data. For each unexpected acoustic signal captured, the color on the graph represents the intensity of that signal. Alternatively, the graph could be represented in monochrome as a contour plot.
These highlighted areas on the graph allow a clinician to distinguish low-energy noise (which may or may not be a sign of non-obstructive CAD) from high-energy noise (which is a likely indicator of obstructive CAD). It is also easy for a clinician to understand the frequency of high-energy noise, as well as the timing of this noise - this is important as different cardiac conditions may be defined by specific frequencies and timings of noise. For example, the data illustrated in FIG. 51 exhibits relatively high intensity noises at approximately t=0.25, 0.75 and 0.8 seconds into the heartbeat.
The system provides a User Interface, and associated navigation, is designed for use on tablets and smartphones, and thus uses common touch-screen user interface paradigms. For example, two fingers moving apart from one another can be used to zoom in to any area on any graph, and two fingers moving together can be used to zoom out. A single finger can be used to highlight any areas on the horizontal axis, and the graph can be zoomed in to that highlighted area by touching a button.
Touching any area of the graph provides information to the user (either in a pop-up window or beside/below the graph) on the values of the horizontal and vertical axes at that point, as well as the“height” information of that point if available (e.g. in the Matrix View). For example, in Stacked Heartbeat View, touching on an individual heartbeat would cause the system to provide the heartbeat number, the maximum intensity of that heartbeat (and the time on the horizontal axis at which the maximum intensity occurs), and the time on the horizontal axis corresponding to the touch. In the case of the Matrix View, touching on any area of the graph would cause the system to provide the frequency, time, and intensity corresponding to the coordinate on the graph that was touched.
For use with a mouse, the user interface paradigms are similar, except with zoom. In this case, zoom can be accomplished through a common desktop/laptop user interface paradigm, such as dedicated zoom in/out buttons in the UI, or mouse wheel scrolling.
For some of these graphs, it is possible to restrict the display of data to some subset of the acoustic sensors (for example, as illustrated in FIG. 47), to combine acoustic data from all sensors (the default mode), or to display data from each sensor individually. This may be helpful to clinicians who want to spatially localize any unexpected acoustic signals. This capability is relevant to the Stacked Heartbeat View, the Bruit Identification View (both modes), and the Bruit Analysis View.
The following features of sound propagation with the body provide for the localization of acoustic sources based upon the relative strengths of the associated acoustic signals from different auscultatory sound sensors 12. First, any sound feature that originates from the heart, acts as a single source for all auscultatory sound sensors 12. For example, the characteristic sound that the heart makes, dub-blub, acts like two sounds coming from two separate locations. Second, for practical purposes, sound travels fast enough in the body that all auscultatory sound sensors 12 would effectively receive each sound individually at substantially the same time. Third, the farther a sound’s origin is from an auscultatory sound sensor 12, the weaker that sound will be when received by that auscultatory sound sensor 12. This is due to the sound energy being dissipated as it travels through the body. In view of these features, using the relative signal strengths from different auscultatory sound sensors 12, the location of the sound source can be triangulated from the locations of the auscultatory sound sensors 12 with the three largest signal strengths, weighted by the relative strengths of those signals, with the resulting calculated location of the sound source being relatively closer to auscultatory sound sensors 12 with relatively stronger signals than to auscultatory sound sensors 12 with relatively weaker signals.
More particularly, referring to FIG. 69, in accordance with one method of localizing the sound source 94 of an auscultatory sound signal 16 based upon associated respective signal strengths PI, P2 and P3 of respective first 12.1, second 12.2 and third 12.3 auscultatory sound sensors that are assumed to be located in Cartesian space on an X-Y plane 96 at respective predetermined sensor locations (XI, Yl), (X2, Y2) and (X3, Y3), at Z=0. Assuming that the distance from each of the auscultatory sound sensors 12.1, 12.2, 12.3 is inversely related to the corresponding associated signal strengths PI, P2 and P3, the corresponding lengths Al, A2, A3 of corresponding edges El, E2, E3 of an associated tetrahedral solid 98 are assumed to be inversely related to the corresponding associated signal strength PI, P2, P3 of the auscultatory sound sensors 12.1, 12.2, 12.3 located at a corresponding vertices VI, V2, V3 of the tetrahedral solid 98 to which the corresponding edge El, E2, E3 is connected. Other than being sufficiently long to meet at a fourth vertex V4 of the tetrahedral solid 98 at location (X, Y, Z), the lengths Al, A2, A3 of the edges El, E2, E3 are otherwise arbitrary, although relatively shorter lengths provide for less geometric dilution of precision (GDOP) of the ultimate determination of the lateral, (X, Y) location of the fourth vertex V4 than to relatively longer lengths. In view of the lateral (X, Y) location of the sound source 94 on the X-Y plane 96 being the same as the lateral (X, Y) location of the fourth vertex V4, the lateral (X, Y) location of the sound source 94 may be determined by solving the following system of three equations in three unknowns (X, Y, Z) for the location (X, Y, Z) of the fourth vertex V4:
(x -x\)2 + (Y -Y\)2 + Z2 = A\2 (31) (X -X2)2 + (Y -Y2)2 + Z2 = A22 (32) (X - X3)2 + (Y - Y3)2 + Z2 = A32 (33)
Following the solution of equations 31-33, the resulting lateral (X, Y) location of the sound source 94 may then be displayed, for example, as a location on a silhouette of a torso, or transformed to a corresponding location on the image of the heart illustrated in FIG. 45.
For some of these graphs, it is possible for the user to directly compare the results of the current test with one or more previous tests. This can be done by simply placing the graphs side- by-side (for example, as illustrated in FIGS. 52 and 53), or by overlaying them on-top of one other, with some variable transparency or color variation so that the clinician can distinguish data from one test versus another. In the case where the graphs are placed on top of one another, the system could fade in to one graph and out from the other slowly, repeatedly in a loop (almost like a video file), so that the clinician can easily see the differences in the processed data between one test and the other. Alternatively, if the graphs are placed on top of one another, a subtraction operation could be performed to highlight only the differences between the two graphs (e.g. red for increase in noise from previous test result, green for decrease). For example, FIG.52 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Spectrogram mode, to confirm success of PCI. Similarly, as another example, FIG. 53 illustrates a comparison of a current test with a previous test, using the Bruit Identification View, Line Graph with Underfill mode, to confirm success of PCI.
The ability to compare the current test with a previous test is critical to the clinicians understanding of the progression of a particular cardiovascular condition (either worsening, or improved as in the case of a PCI procedure). This capability is relevant to the Stacked Heartbeat View, the Bruit Identification View (both modes), and the Bruit Analysis View.
The graphs may be rendered locally or remotely (server-based) or both, depending on the capabilities desired by the clinician and the organization to which he belongs. In most use cases (tablet, phone, desktop, or laptop), the graph rendering will be done locally, either through a web- browser (full or embedded) on the client, or through a graphics library optimized for each specific supported client platform.
In other cases, the rendering may be done on the server side - graphs may be generated and exported to JPEG (or other similar) format so that they can be emailed or sent via instant message to interested parties.
While specific embodiments have been described in detail in the foregoing detailed description and illustrated in the accompanying drawings, those with ordinary skill in the art will appreciate that various modifications and alternatives to those details could be developed in light of the overall teachings of the disclosure. It should be understood, that any reference herein to the term“or” is intended to mean an“inclusive or” or what is also known as a“logical OR”, wherein when used as a logic statement, the expression“A or B” is true if either A or B is true, or if both A and B are true, and when used as a list of elements, the expression“A, B or C” is intended to include all combinations of the elements recited in the expression, for example, any of the elements selected from the group consisting of A, B, C, (A, B), (A, C), (B, C), and (A, B, C); and so on if additional elements are listed. Furthermore, it should also be understood that the indefinite articles "a" or "an", and the corresponding associated definite articles“the’ or“said”, are each intended to mean one or more unless otherwise stated, implied, or physically impossible. Yet further, it should be understood that the expressions“at least one of A and B, etc.”,“at least one of A or B, etc.”, “selected from A and B, etc.” and“selected from A or B, etc.” are each intended to mean either any recited element individually or any combination of two or more elements, for example, any of the elements from the group consisting of“A”,“B”, and“A AND B together”, etc. Yet further, it should be understood that the expressions“one of A and B, etc.” and“one of A or B, etc.” are each intended to mean any of the recited elements individually alone, for example, either A alone or B alone, etc., but not A AND B together. Furthermore, it should also be understood that unless indicated otherwise or unless physically impossible, that the above-described embodiments and aspects can be used in combination with one another and are not mutually exclusive. Accordingly, the particular arrangements disclosed are meant to be illustrative only and not limiting as to the scope of the invention, which is to be given the full breadth of the appended claims, and any and all equivalents thereof.
What is claimed is:

Claims

1. A method of segmenting an auscultatory sound signal, comprising:
a. receiving an electrographic signal from an ECG sensor;
b. generating an electrographic envelope signal representing an envelope responsive to an even power of said electrographic signal;
c. locating a plurality of peaks of said electrographic envelope signal corresponding to a corresponding plurality of R-peaks of said electrographic signal, wherein the operation of locating said plurality of peaks of said electrographic envelope signal comprises: i. determining a plurality of maximum values of said electrographic envelope signal, and a corresponding associated plurality of times of occurrence, for each of a plurality of temporal windows that are separated in time from an adjacent one another by a hop period;
ii. determining from said plurality of maximum values of said electrographic envelope signal, at said corresponding associated plurality of times of occurrence, a set of maximum values and corresponding times for which each value of said set of maximum values is less than a first threshold and greater than a second threshold; iii. determining a plurality of heart rates responsive to said corresponding times associated with each adjacent-in-time pair of maximum values of said set of maximum values;
iv. selecting from said each adjacent-in-time pair of maximum values of said set of maximum values at least one adjacent-in-time pair of maximum values of said set of maximum values for which a heart rate of said plurality of heart rates is greater than a first fraction less than unity times a median value of said plurality of heart rates and for which said heart rate of said plurality of heart rates is less than a first factor greater than unity times said median value of said plurality of heart rates, wherein the operation of selecting from said each adjacent-in-time pair of maximum values of said set of maximum values provides for locating said plurality of peaks of said electrographic envelope signal corresponding to said corresponding plurality of R- peaks of said electrographic signal;
d. receiving at least one auscultatory sound signal from a corresponding at least one auscultatory sound-or-vibration sensor; and
e. segmenting said at least one auscultatory sound signal into at least one heart-cycle segment responsive to said plurality of peaks of said electrographic envelope signal.
2. A method of segmenting an auscultatory sound signal as recited in claim 1, further comprising for each one said at least one heart-cycle segment:
a. generating an auscultatory envelope signal representing an envelope responsive to an even power of said auscultatory sound signal within said one said at least one heart-cycle segment;
b. locating first and second peaks of said auscultatory envelope signal corresponding to respective first and second heart sounds;
c. generating a local mathematical model of at least said second peak of said auscultatory envelope signal, wherein said local mathematical model comprises a quadratic equation; and
d. locating first and second roots of said local mathematical model so as to provide for determining an associated start of diastole of said one said at least one heart-cycle segment.
3. A method of segmenting an auscultatory sound signal as recited in claim 2 further comprising filtering said electrographic signal with a low pass filter prior the operation of generating said electrographic envelope signal.
4. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising normalizing said electrographic signal prior the operation of generating said electrographic envelope signal.
5. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein each sample of said electrographic envelope signal comprises a sum of products within a corresponding sliding window of time associated with said sample of said electrographic envelope signal, each product of said sum of products comprises an element multiplied by a natural logarithm of said element, wherein said element comprises a value of a corresponding sample of said electrographic signal raised to a fourth power, and said corresponding sample of said electrographic signal is within said sliding window of time.
6. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising validating each peak of said plurality of peaks of said electrographic envelope signal, wherein the operation of validating said plurality of peaks of said electrographic envelope signal comprises at least one comparison operation selected from the group consisting of comparing a duration of time between said peak and a next subsequent peak with a first threshold, and comparing an absolute magnitude difference between a magnitude of said peak and a median value of corresponding magnitudes of said plurality of peaks with a second threshold; and responsive to said at least one comparison operation, determining whether or not to ignore said peak for purposes of subsequent operations.
7. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising filtering said at least one auscultatory sound signal with a low pass filter prior the operation of generating said auscultatory envelope signal.
8. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising filtering said at least one auscultatory sound signal with a Savitzky-Golay filter prior the operation of generating said auscultatory envelope signal.
9. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein each sample of said auscultatory envelope signal comprises a sum of products within a corresponding sliding window of time associated with said sample of said auscultatory envelope signal, each product of said sum of products comprises an element multiplied by a natural logarithm of said element, wherein said element comprises a value of a corresponding sample of said auscultatory sound signal raised to a second power, and said corresponding sample of said auscultatory sound signal is within said sliding window of time.
10. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising smoothing said auscultatory envelope signal prior to the operation of locating said first and second peaks of said auscultatory envelope signal, wherein the operation of smoothing said auscultatory envelope signal comprises filtering said auscultatory envelope signal with a filter selected from the group consisting of a moving average filter and a
Savitzky-Golay smoothing filter.
11. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein the operation of locating said first and second peaks of said auscultatory envelope signal comprises:
a. comparing said auscultatory envelope signal with an adjustable threshold;
b. locating peak values of said auscultatory envelope signal in excess of said adjustable threshold; and
c. adjusting said adjustable threshold until only two said peak values of said auscultatory envelope signal are located in excess of said adjustable threshold for said one said at least one heart-cycle segment.
12. A method of segmenting an auscultatory sound signal as recited in claim 2, further comprising validating said first and second peaks of said auscultatory envelope signal responsive to a comparison of a duration of time between said first and second peaks of said auscultatory envelope signal in comparison with a threshold.
13. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein the operation of generating said local mathematical model comprises: a. selecting a pair of points of said auscultatory envelope signal spanning in time a peak of said auscultatory envelope signal selected from the group consisting of said first peak of said auscultatory envelope signal and said second peak of said auscultatory envelope signal, wherein each point of said pair of points of said auscultatory envelope signal is selected so that a corresponding associated magnitude of auscultatory envelope signal is not in excess of a fraction of a magnitude of said peak of said auscultatory envelope signal; and
b. mathematically fitting a quadratic function of time to a set of points comprising said pair of points of said auscultatory envelope signal and to said peak of said auscultatory envelope signal.
14. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein said associated start of diastole is associated with a later-occurring root of said first and second roots of said local mathematical model of said second peak of said auscultatory envelope signal.
15. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein a region of diastole extends from said associated start of diastole until an end of said one said at least one heart-cycle segment, further comprising analyzing said region of diastole of said auscultatory sound signal associated with said one said at least one heart-cycle segment to identify whether or not one or more outliers are present in said auscultatory sound signal.
16. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein a region of diastole extends from said associated start of diastole until an end of said one said at least one heart-cycle segment, further comprising analyzing said region of diastole of said auscultatory sound signal associated with said one said at least one heart-cycle segment to identify whether or not said auscultatory sound signal exceeds one or more noise thresholds.
17. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein a region of diastole extends from said associated start of diastole until an end of said one said at least one heart-cycle segment, further comprising normalizing said region of diastole of said auscultatory sound signal with respect to a magnitude of said second peak of said auscultatory envelope signal.
18. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein a region of diastole extends from said associated start of diastole until an end of said one said at least one heart-cycle segment, further comprising:
a. calculating a maximum value of a plurality of mean power levels, wherein each mean power level of said plurality of mean power levels corresponds to a power level of said auscultatory sound signal within a sliding window of a plurality of sliding windows within said region of diastole of said auscultatory sound signal, and different mean power levels of said plurality of mean power levels correspond to different sliding windows of said plurality of sliding windows; and
b. comparing said maximum value of said plurality of mean power levels with a threshold to determine whether or not to subsequently process said one said at least one heart-cycle segment.
19. A method of segmenting an auscultatory sound signal as recited in claim 2, wherein a region of diastole extends from said associated start of diastole until an end of said one said at least one heart-cycle segment, and said at least one heart-cycle segment comprises a plurality of heart-cycle segments, further comprising synchronizing each of said plurality of heart-cycle segments with respect to said associated start of diastole.
20. A method of segmenting an auscultatory sound signal as recited in claim 19, further comprising normalizing an associated time period of each said region of diastole of said plurality of heart-cycle segments with respect to one of said plurality of heart-cycle segments.
21. A method of segmenting an auscultatory sound signal as recited in claim 20, wherein said one of said plurality of heart-cycle segments is the slowest of said plurality of heart-cycle segments.
22. A method of segmenting an auscultatory sound signal as recited in claim 20, further comprising interpolating one or more of said plurality of heart-cycle segments so as to provide for a uniform effective sampling rate for each of said plurality of heart-cycle segments.
23. A method of segmenting an auscultatory sound signal as recited in claim 1, wherein a duration of each temporal window of said plurality of temporal windows is sufficient to span a single heart cycle.
24. A method of segmenting an auscultatory sound signal as recited in claim 1, wherein said first threshold is a second factor in excess of unity times a median value of said plurality of maximum values of said electrographic envelope signal over a period of time spanning a plurality of heat cycles, and a value of said second factor is sufficient so that the values of at least one pair of said set of maximum values associated with at least one heart cycle of said plurality of heat cycles are each less than said first threshold.
25. A method of segmenting an auscultatory sound signal as recited in claim 1, wherein said second threshold is a second fraction less then unity times a median value of said plurality of maximum values of said electrographic envelope signal, and a value of said second fraction is sufficient so that the values of at least one pair of said set of maximum values associated with at least one heart cycle of said plurality of heat cycles are greater than said second threshold.
26. A method of segmenting an auscultatory sound signal as recited in claim 1, wherein the operation of selecting from said each adjacent-in-time pair of maximum values of said set of maximum values is bypassed if a time difference between said corresponding times associated with said adjacent-in-time pair of maximum values is less than a third threshold.
27. A method of segmenting an auscultatory sound signal as recited in claim 26, wherein said third threshold is in excess of a heart-cycle period corresponding to a maximum expected heart rate.
28. A method of segmenting an auscultatory sound signal as recited in claim 1, wherein said first fraction is about 0.6 and said first factor is about 1.4.
29. A method of segmenting an auscultatory sound signal, comprising:
a. receiving an electrographic signal from an ECG sensor;
b. generating an electrographic envelope signal representing an envelope responsive to an even power of said electrographic signal;
c. locating a plurality of peaks of said electrographic envelope signal corresponding to a corresponding plurality of R-peaks of said electrographic signal;
d. receiving at least one auscultatory sound signal from a corresponding at least one auscultatory sound-or-vibration sensor;
e. segmenting said at least one auscultatory sound signal into at least one heart-cycle segment responsive to said plurality of peaks of said electrographic envelope signal; and f. for at least one said at least one heart-cycle segment:
i. identifying a temporal region within each said at least one said at least one heart- cycle segment, wherein a location of said temporal region is responsive to a duration of said at least one said at least one heart-cycle segment; and ii. determining at least one swing value responsive to a difference between maximum and minimum amplitude values of said auscultatory sound signal within at least one temporal segment of said temporal region.
30. A method of segmenting an auscultatory sound signal as recited in claim 29, further comprising utilizing said at least one first swing value as a first feature input set to a classifier to provide for determining whether or not a patient associated with said auscultatory sound signal likely exhibits coronary -artery disease.
31. A method of segmenting an auscultatory sound signal as recited in claim 30, wherein said temporal region is associated with at least one sound selected from the group consisting of an S3 sound, an S4 sound and an S5 sound.
32. A method of segmenting an auscultatory sound signal as recited in claim 31, wherein said temporal region is located relative to a start of diastasis within diastole of said heart-cycle segment.
33. A method of segmenting an auscultatory sound signal as recited in claim 32, wherein said start of diastasis is assumed to occur at a temporal offset from a start of said heart-cycle segment, with said temporal offset being given in milliseconds by 350 plus 0.3 times the quantity (TRR-350), wherein TRR is the duration of said heart-cycle segment in milliseconds.
34. A method of segmenting an auscultatory sound signal as recited in claim 32, wherein said temporal region either spans said start of diastasis or terminates at said start of diastasis.
35. A method of segmenting an auscultatory sound signal as recited in claim 32, wherein a duration of said temporal region is between 80 and 120 milliseconds.
36. A method of segmenting an auscultatory sound signal as recited in claim 32, further comprising:
a. generating a Short-Time Fourier Transform of said auscultatory sound signal within said temporal region; and
b. utilizing at least one frequency-domain feature from or responsive to said Short-Time
Fourier Transform of said auscultatory sound signal within said temporal region as a second feature input set to said classifier to provide for determining whether or not said patient associated with said auscultatory sound signal likely exhibits coronary-artery disease.
37. A method of segmenting an auscultatory sound signal as recited in claim 31, wherein said temporal region is located at the end of said heart-cycle segment.
38. A method of segmenting an auscultatory sound signal as recited in claim 37, wherein a duration of said temporal region is between 10 and 25 percent of the duration of said heart- cycle segment.
39. A method of segmenting an auscultatory sound signal as recited in claim 37, further comprising prior to the operation of determining said at least one first swing value, generating a filtered auscultatory sound signal by filtering said auscultatory sound signal within said heart-cycle segment using a filter having a cutoff frequency the provides for passing either an audible range of frequencies or passing an inaudible range of frequencies below said audible range of frequencies, wherein said at least one first swing value is determined responsive to said filtered auscultatory sound signal.
40. A method of segmenting an auscultatory sound signal as recited in claim 37, further comprising:
a. determining a second swing value responsive to a difference between maximum and minimum amplitude values of said auscultatory sound signal during an S2 sound interval at the beginning of diastole of said heart-cycle segment;
b. determining a ratio of said first swing value divided by said second swing value as a measure of a likelihood that said patient exhibits said S4 sound; and
c. utilizing said ratio of said first swing value divided by said second swing value as a third feature input set to said classifier to provide for determining whether or not a patient associated with said auscultatory sound signal likely exhibits coronary-artery disease.
PCT/US2020/029970 2017-10-21 2020-04-24 Medical decision support system WO2020219991A1 (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CA3137910A CA3137910A1 (en) 2019-04-24 2020-04-24 Medical decision support system
EP20794539.5A EP3958743A1 (en) 2019-04-24 2020-04-24 Medical decision support system
US17/509,018 US20220061797A1 (en) 2017-10-21 2021-10-24 Medical decision support system

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US201962838270P 2019-04-24 2019-04-24
US201962838296P 2019-04-24 2019-04-24
US62/838,296 2019-04-24
US62/838,270 2019-04-24

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
US16/854,894 Continuation-In-Part US11284827B2 (en) 2017-10-21 2020-04-21 Medical decision support system

Related Child Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2021/044854 Continuation-In-Part WO2022032041A1 (en) 2017-10-21 2021-08-05 Medical decision support system

Publications (1)

Publication Number Publication Date
WO2020219991A1 true WO2020219991A1 (en) 2020-10-29

Family

ID=72941420

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2020/029970 WO2020219991A1 (en) 2017-10-21 2020-04-24 Medical decision support system

Country Status (3)

Country Link
EP (1) EP3958743A1 (en)
CA (1) CA3137910A1 (en)
WO (1) WO2020219991A1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022032041A1 (en) * 2020-08-05 2022-02-10 Ausculsciences, Inc. Medical decision support system
US11284827B2 (en) 2017-10-21 2022-03-29 Ausculsciences, Inc. Medical decision support system
CN114343642A (en) * 2021-10-22 2022-04-15 厦门大学 Fatigue driving detection method and system based on heart rate variability index

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040267148A1 (en) * 2003-06-27 2004-12-30 Patricia Arand Method and system for detection of heart sounds
US20100094152A1 (en) * 2006-09-22 2010-04-15 John Semmlow System and method for acoustic detection of coronary artery disease
US20120130263A1 (en) * 2009-08-03 2012-05-24 Diacoustic Medical Devices (Pty) Ltd Medical decision support system
US20150238147A1 (en) * 2002-03-18 2015-08-27 Sonomedica, Inc. Method of processing cardiovascular sound signals

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150238147A1 (en) * 2002-03-18 2015-08-27 Sonomedica, Inc. Method of processing cardiovascular sound signals
US20040267148A1 (en) * 2003-06-27 2004-12-30 Patricia Arand Method and system for detection of heart sounds
US20100094152A1 (en) * 2006-09-22 2010-04-15 John Semmlow System and method for acoustic detection of coronary artery disease
US20120130263A1 (en) * 2009-08-03 2012-05-24 Diacoustic Medical Devices (Pty) Ltd Medical decision support system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
GIORDANO, NOEMI, KNAFLITZ, MARCO: "A Novel Method for Measuring the Timing of Heart Sound Components through Digital Phonocardiography", SENSORS 2019, vol. 19, 19 April 2019 (2019-04-19), pages 1 - 16, XP055755526, Retrieved from the Internet <URL:https://www.mdpi.com/1424-8220/19/8/1868> [retrieved on 20200706], DOI: 10.3390/s19081868 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11284827B2 (en) 2017-10-21 2022-03-29 Ausculsciences, Inc. Medical decision support system
US11896380B2 (en) 2017-10-21 2024-02-13 Ausculsciences, Inc. Medical decision support system
WO2022032041A1 (en) * 2020-08-05 2022-02-10 Ausculsciences, Inc. Medical decision support system
CN114343642A (en) * 2021-10-22 2022-04-15 厦门大学 Fatigue driving detection method and system based on heart rate variability index

Also Published As

Publication number Publication date
CA3137910A1 (en) 2020-10-29
EP3958743A1 (en) 2022-03-02

Similar Documents

Publication Publication Date Title
US11896380B2 (en) Medical decision support system
Ha et al. Contactless seismocardiography via deep learning radars
Ansari et al. A review of automated methods for detection of myocardial ischemia and infarction using electrocardiogram and electronic health records
Martis et al. Current methods in electrocardiogram characterization
EP2375982B1 (en) Method and apparatus for the analysis of ballistocardiogram signals
EP3048972B1 (en) Processing apparatus, processing method and system for processing a physiological signal
US11311201B2 (en) Feature selection for cardiac arrhythmia classification and screening
Castro et al. Heart sound segmentation of pediatric auscultations using wavelet analysis
JP6457117B2 (en) Continuous non-invasive monitoring of pregnant subjects
Geweid et al. A new automatic identification method of heart failure using improved support vector machine based on duality optimization technique
EP3958743A1 (en) Medical decision support system
WO2019079829A9 (en) Method of preprocessing and screening auscultatory sound signals
JP2018512243A5 (en)
CN107072550A (en) Body moves recording method and device
Wang et al. Temporal-framing adaptive network for heart sound segmentation without prior knowledge of state duration
Barschdorff et al. Automatic phonocardiogram signal analysis in infants based on wavelet transforms and artificial neural networks
Albaba et al. Assessing the signal quality of electrocardiograms from varied acquisition sources: A generic machine learning pipeline for model generation
Singh et al. Short and noisy electrocardiogram classification based on deep learning
CN111031902B (en) Multi-sensor stroke output monitoring system and analysis method
Phinyomark et al. Applications of variance fractal dimension: A survey
Oliveira et al. Coupled hidden Markov model for automatic ECG and PCG segmentation
WO2022032041A1 (en) Medical decision support system
Noponen et al. Invariant trajectory classification of dynamical systems with a case study on ECG
US20220061797A1 (en) Medical decision support system
Roy et al. Reconstruction of corrupted and lost segments from photoplethysmographic data using recurrent neural network

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20794539

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 3137910

Country of ref document: CA

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2020794539

Country of ref document: EP

Effective date: 20211124