FI129789B

FI129789B - A method, an apparatus and a computer program product for determination of pulse transit time

Info

Publication number: FI129789B
Application number: FI20205775A
Authority: FI
Inventors: Ilkka Korhonen; Maunu Toiviainen
Original assignee: Firstbeat Analytics Oy
Priority date: 2020-07-29
Filing date: 2020-07-29
Publication date: 2022-08-31
Also published as: FI20205775A1

Abstract

The invention relates to a method, an apparatus and a software program product for determining a pulse transit time, comprising detecting first and second activity of a heart rate of a user, determining sequences of the first and the second activity respectively from the first and the second activity of the heart rate as a function of time, identifying sequences of the first activity and sequences of the second activity corresponding to the same heart beat, calculating the average mutual offset of the sequences of the first activity and sequences of the second activity, calculating the change in the offsets of the sequences of the first activity and sequences of the second activity, and determining the changes in the pulse transit time, PTT as a difference between timely matched sequences of the first activity and sequences of the second activity.

Description

A method, an apparatus and a computer program product for determination of pulse transit time Technical field

The application relates to a method, an apparatus and a software product for determining a pulse transit time.

Background

A pulse transit time is a time it takes for a pulse wave to transit between two arterial sites.

A pulse transit time, PTT, may be determined for example by using an electrocardiogram, ECG and a photoplethysmogram, PPG.

ECG determines electrical activity of the heart in order to indicate the initiation of the mechanical pulse from the heart to the aorta.

PPG measures arrival of the blood volume pulse to the peripheral circulation at the point of measurement.

Alternatively, PTT may be determined from two sensors measuring progression of the blood volume pulse from two distinct locations, for example an upper arm and a wrist or a finger.

The mechanical pulse may be measured with PPG or, for example, from a pressure sensor.

PTT is then determined as a time difference between the two signals, i.e. a time difference between initiation of the pulse from the heart and arrival of the pulse at the peripheral circulation at two distinct locations.

ECG, PPG or pressure sensing measurement device may be held somewhere on a body of a user.

PTT is typically in the range between 50 ms and 150 ms, depending on at least the N placement of the sensor devices and physiological factors, such as blood N pressure.

PPG is typically detected at a device attached at a peripheral area S of the body of the user, like a wrist, a finger or an ear.

Instead of two separate & devices it is possible to use a single device comprising two sensors.

Ek 30 so In order to determine the pulse transit time, the devices or components N measuring the two signals, e.g. of an ECG and a PPG; two distinct PPG S sensors; or a PPG and a pressure sensor, or any combination thereof, need to be carefully synchronized.

This may require specific pre-synchronization of used devices or components.

This may include, for example, synchronization of a signal sampling circuit or a processing circuit or dedicated timers of such,

like crystal oscillators; or calibrating the signals by measuring the time and frequency differences between the device clocks and compensating the difference computationally in the signal processing circuitry processing PPG or ECG signals.

Summary An aim is to determine a pulse transit time, PTT, using two independent measurements, optionally originating from different sources, components or devices, without a pre-synchronization of the sources, components or devices. The pulse transit time, PTT, may be determined using two independent measurements of a first activity of a heart of a user and a second activity of the heart of the user. The first activity of the heart may be measured from the electrical activity of the heart or from the mechanical activity of the heart. The second activity of the heart may be measured from the mechanical activity of the heart. The mechanical activity of the heart may be measured from blood flow pulsation of the user. The measured results of the two independent measurements are synchronized using a detected biological rhythm of the user. This enables synchronizing measured data of two independent measurements without a need to synchronize the measuring devices before the measurements. According to an aspect of the invention a method for determining a pulse transit time comprises an arrangement to identify the two independent measurements belonging to the same heart beat and calculating their artefact-free unbiased N time difference as the pulse transit time PTT by detecting first and second N activity of a heart rate of a user, removing artefacts from the detected first and S second activity of a heart rate, determining seguences of the first and the & second activity from the first activity of the heart rate as function of time, I 30 thereby identifying sequences of the first and the second activity so corresponding to the same heart beat (contraction time). The method further N comprises calculating the average mutual offset of the seguences of the first S and the second activity, calculating the change in the offsets of the sequences of the first and the second activity and correcting the sequences for any errors in synchronization. Finally, the method determines the changes in the pulse transit time, PTT as a difference between timely matched sequences of the first activity and sequences of the second activity. According to another aspect of the invention an apparatus for determining a pulse transit time comprises at least one first sensor, at least one second sensor, a microprocessor, a memory, a storage device, and a software product configuration. The at least one first sensor detects electrical or mechanical activity of a heart rate of a user and the at least one second sensor detects mechanical activity of the heart rate of the user. The apparatus also comprises an arrangement to detect artefacts from the detected first and second activity of a heart rate. The detected artefacts may be removed, ignored in the analysis or replaced with interpolated values. The apparatus further comprises an arrangement to determine sequences of the first and the second activity of the heart rate as function of time, an arrangement to identify sequences of the first and the second activity corresponding to the same heart beat (contraction time), an arrangement to calculate the average mutual offset of the sequences of the first and the second activity, an arrangement to calculate the change in the offsets of the sequences of the first and the second activity and an arrangement to determine the changes in the pulse transit time, PTT as a difference between timely matched sequences of the first activity and sequences of the second activity.

According to a further aspect of the invention a software program product for determining a pulse transit time comprises instructions, which when executed carry out at least one of the steps described below. The software program N product retrieves detected and/or stored first and second activity of a heart rate N into memory and process the detected and/or the stored first and second S activity of a heart rate in order to remove artefacts from the detected and/or Q the stored first and second activity of a heart rate. The software program z 30 product then determines seguences of the first and the second activity of the so heart rate as function of time and identify sequences of the first and the second N activity corresponding to the same heart beat (contraction time). The software S program product further calculates the average mutual offset of the sequences of the first and the second activity and additionally calculate the change in the — offsets of the seguences of the first and the second activity. The software program product determines the changes in the pulse transit time, PTT as a difference between timely matched sequences of the first activity and sequences of the second activity, store the calculated pulse transit time, PTT and output the pulse transit time, PTT to the user.

Brief description of the drawings Figure 1 illustrates, by way of an example, a method for measuring pulse transit time.

Figure 2 illustrates, by way of an example, an apparatus arranged to measure pulse transit time. Figure 3A illustrates, by way of an example, an apparatus arranged to measure pulse transit time. Figure 3B illustrates, by way of an example, an apparatus arranged to measure pulse transit time. Figure 4A illustrates, by way of an example, phases of an embodiment for providing pulse transit time. Figure 4B illustrates, by way of an example, phases of an embodiment for providing pulse transit time.

N Figure 5 illustrates the detection of missing or extra pulses in the N heart rate timeseries

S & The Figures are illustrating aspects of the invention and aid in understanding I 30 relating subject-matter and context. a

LO R Detailed description

S

O N A pulse transit time, PTT refers to a time of a pulse pressure waveform to propagate through the arterial tree, between the two arterial sites. The pulse pressure waveform illustrates ejection of blood from left ventricle, which moves with velocity greater than forward movement velocity of blood itself. The pulse transit time is inversely proportional to blood pressure. PTT for measuring blood pressure enables easy, reliable, comprehensive and long-term monitoring method to, for example to measure physical and mental stress 5 during a longer period of time. Current solutions for continuous PTT analysis lack sufficient real-time synchronization accuracy. In this application, a novel method, apparatus and software for accurate PTT determination without pre- synchronization is presented.

Two independent measurements of activity need to be synchronized in order to enable determination of the pulse transit time, PTT. The vast majority of measuring devices include a clock that is at least sufficiently well synchronized. When two or more independent devices are used, however, the independent devices are likely unsynchronized due to technological drift. Absolute PTT — values cannot be measured without proper synchronization of the two devices. Therefore it is not possible to measure blood pressure variability, BPV, as a result of the synchronization drift between the two independent devices unless the synchronization drift is compensated for. The proposed method may be used to calculate, how much the PTT has changed (in milliseconds) relative to a selected reference point.

Figure 1 illustrates, by way of an example, a method for measuring a pulse transmit time, PTT. A pulse transit time, PTT may be determined from simultaneously measured heart rate from more than one location of a human body. In order to measure the transit time, the locations should have a different N distance to the heart, preferably at least one location being closer or located N proximally to the heart and at least one location being farther or located distally S to the heart. A first set of heart rate data is collected at phase 101. A second & set of heart rate data is collected at phase 102. The first set of heart rate data I 30 and the second set of heart rate data are measured simultaneously, or at least so partly parallel. PTT may be calculated, at phase 106 of Figure 1, as a temporal N difference between heart beat timestamps at more than one simultaneously S measured locations.

O

N At least one or both heart rate measurements for PTT are mechanical. PTT may be measured from electrical activity of a heart of auser and a mechanical activity of blood of the user. The first heart beat measurement may be electrical or mechanical. The second heart beat measurement may mechanical. At least part of the first and the second heart beat measurements must correspond to the same heart beats. Electrical activity of a heart may be measured using electrocardiogram, ECG. Electrical activity measurement may provide a cardiac cycle length. Optionally, the calculated PTT may be outputted to the user at phase 107. Cardiac cycle lengths may be determined as RR intervals, RRI, where R is a point corresponding to a peak of the ECG wave and RR is the interval between successive peaks or Rs. Cardiac cycle lengths may be determined as PP intervals, PPI, where P is a point corresponding to a peak of the photoplethysmogram, PPG, wave and PP is the interval between successive peaks or Ps. Cardiac cycle lengths or pulsation intervals may correspond to — inter-beat intervals. Electric or mechanical first inter-beat intervals may be called IBI1. Mechanical second inter-beat intervals may be called IBI2. PTT may be determined from detecting the IBls from the first activity of a heart, IBI1, and detecting IBIs from the second activity of the heart, IBI2, from more than one location simultaneously. The measured IBI1 and IBI2 values may be processed at the PTT apparatus in order to provide PTT, which may be outputted, as illustrated in Figure 2. Method may comprise correcting missing values, as illustrated at phase 103 of Figure 1. Missing values in the seguences of first activity of a heart, and in the sequences of second activity of the heart may be compensated by adding N one or more values to compensate for the missing value. The added value(s) N may be arranged to be identified as artefacts. In addition to missing values, S the ECG and/or PPG signals may contain extra peaks which are erroneously & detected as heart beats. The artefact detection and removal algorithm should I 30 also be able to detect these extra peaks which result in inter-beat intervals that so are shorter than the surrounding non-artefact inter-beat intervals.

N S Method may comprise removing artefacts, as illustrated at phase 104 of Figure N 1. Artefacts may be detected and removed from the sequences of first activity ofaheart, and from the sequences of second activity of the heart. Alternatively, the detected artefacts may be merely ignored in the analysis. Artefact detection may be based on artefact detection algorithms. Artefact detection may contain outlier removal. However, it is also possible that the heart beat timestamps classified as artefacts are corrected. Missing beats result in overly long inter-beat intervals. These long inter-beat intervals may be replaced with realistic interpolated values (based on the surrounding heart beat timestamps that were classified as non-artefacts). Extra beats (e.g., additional peaks in the pulse wave that are erroneously detected as heart beats) usually result in IBls that are much shorter than the surrounding non-artefact IBIs. The short IBls may be combined (summed) into one IBI whose duration is close to the values of the surrounding non-artefact IBls. Artefact correction is needed if the heart beat timestamp or IBI sequence is re-sampled into an evenly-spaced time series. The method of Figure 1 comprises matching the collected heart rates based on the heart beat at phase 105. 1-100 consecutive cardiac cycle lengths may be identified as cardiac cycle sub-seguence. Sub-seguences from IBI1 and IBI2 may be identified belonging to the same individual heart beats. Sub- seguences may be called templates. Timeseries of more than one measured seguences may be shifted in temporal dimension to align individual heart beats based on the similarity of sub-seguences. Each sub-seguence may be shifted individually. Shifting of the sub-seguences in time may be called template matching. Individual timeseries may be described as a cumulative sum of the seguences. Average mutual offset of the cardiac cycle length sequences and the changes N in the offset of the cardiac cycle length seguences may be calculated from N changes in the pulse transit time as a difference between timely matched sub- S sequences of the first activity of the heart rate, IBI1 and the second activity of S the heart rate, IBI2. I 30 so Identification of the sequences belonging to the same heart beat may be N determined by template matching the sub-seguences of the first activity of the S heart rate with the sub-sequences of the second activity of the heart rate.

O N

Sequences of the first activity of a heart IBI1 may be determined via continuous measurement of cardiac cycle lengths, RRIs, or via continuous measurement of pulsations of blood volume and/or pressure peripheral to the heart, PPIs.

Sequences of the second activity of a heart IBI2 may be determined via continuous measurement of pulsations of blood volume and/or pressure peripheral to the heart or IBls.

Determining sequences of first activity of a heart, IBI1 as a function of time and determining the sequences of second activity of the heart, IBI2 as a function of time may include generation of a time axis for the sequences of the IBl1s from the first activity and generation of a time axis for the seguences of the IBI2s from the second activity.

Determining sequences of first activity of a heart, IBI1 as a function of time and determining the sequences of second activity of the heart, IBI2 as a function of time may include generation of a time axis by a cumulative sum of the sequences of the IBI1s of the first activity and generation of a time axis by a cumulative sum of the sequences of the IBI2s of the second activity.

Determining seguences of first activity of a heart as a function of time and determining the seguences of second activity of the heart as a function of time may include calculating time vectors for the first activity of the heart and for the second activity of the heart.

N Determining sequences of first activity of a heart, IBI1 as a function of time and N determining the sequences of second activity of the heart, IBI2 as a function S of time may include calculating time vectors for the sequences of the IBI1s of & the first activity of the heart and of the IBI2s of the second activity of the heart, I 30 optionally by cumulative sums, or other methods preserving temporal order so and length of the time vectors.

N S The sequences of first activity and the sequences of second activity of a heart may be matched with each other based on time via linear or non-linear fitting, optionally based, at least partly, on the internal clock offsets between a sensor arranged to measure the first activity of the heart and a sensor arranged to measure the second activity of the heart.

Method of Figure 1 comprises matching heart rates to the same heart beat at phase 105. Matching may be done for sub-sequences of the measured heart rates. Matching may include identifying starting locations of the sub- sequences. Starting locations of sub-sequences of first activity of a heart as a function of time, and of sub-sequences of second activity of the heart as a function of time may identified from the respective first and second activity sequences, optionally the sub-sequences including a predetermined length or a predetermined number of elements, for example 1-100, preferably 10-50 elements.

Each sub-sequence of first activity of a heart may be matched to a sub- sequence of second activity of the heart. Offset between the sub-sequences may be calculated at the matching point of time.

A matching sub-sequence of second activity of a heart for each sub-sequence — of first activity of the heart may be identified via error metrics, optionally via minimum mean absolute percentage error.

One or more sub-sequences of second activity of a heart may be shifted, stretched or contracted as a function of time based on the sub-sequences of first activity of the heart. Similarly, one or more sub-sequences of first activity N of a heart may be shifted, stretched or contracted as a function of time based N on the sub-seguences of second activity of a heart.

S & Figures 3A and 3B illustrate, by way of example, an apparatus arranged to I 30 measure PTT. A sensor or a sensor device refers to a sensor enabling the so measurement of the electrical activity of the heart and/or a sensor enabling the N measurement of the mechanical activity of the heart. Sensor for detecting the S mechanical activity of the heart may include the sensor detecting the electrical activity of the heart. Alternatively, each or some of the sensors may be individual sensors. The apparatus may consist of one or more sensors recording the electrical activity of a heart. Electrical activity of the heart may be recorded as IBI1. The apparatus may consist of one or more sensors recording the mechanical activity of the heart as inter-beat intervals, IBI2s. Consecutive IBl1s and IBl2s form timeseries, which may include for example sub- seguences and seguences of electrical and/or mechanical activity of the heart. The sensors for mechanical activity of the heart may be positioned peripheral to the sensors measuring the electrical activity of the heart. One or more of the sensors may be contact sensors. One or more of the sensors may be remote sensing sensors.

The sensor for recording the electrical activity of the heart may be measured using a first sensor S1 of Figures 3A and 3B. The fist sensor S1 may comprise, for example, a sensor measuring electrocardiogram. Electrocardiogram may be measured as heart rate R-peak intervals, RRIs. Mechanical activity of the heart measured from blood flow pulsation of the user may be measured using a second sensor S2 of Figures 3A and 3B. The second sensor S2 may comprise a photoplethysmogram, PPG, a pressure sensor, a ballistogardiogram, contact ballistocardiography, a non-contact red or infrared wavelength sensing camera or alike blood volume induced pressure or blood volume change sensing device or component. Mechanical activity measurement of the second sensor S2 may provide pulsation of a blood pressure, or volume in peripheral area. Mechanical activity may be measured as pulsation inter-beat intervals, IBls. Typically, PPG may be used to measure heart beat pulse at distant, peripheral point to a heart while ECG may be used to get near real-time electrical signal of the heart beat from close to the heart.

Alternatively, for example, two PPGs having a different distance to the heart N may be used. Therefore, the first sensor 51 and the second sensor S2 may N both be mechanical sensors.

S & An apparatus of Figure 3A for determining a pulse transit time may be an I 30 electronic device comprising at least one processor uP, and at least one so memory MEM or a storage device. The storage device may be arranged to N store measured heart rate values and executable instructions for providing S PTT based on two simultaneously measured sets of heart rate data. The at least one processor may be arranged to execute the instructions in order to cause the PTT to be calculated based on the two simultaneously measured sets of heart rate data.

The set of heart rate data collected via second sensor S2 may be received via adapter AD2. The heart set of rate data collected via first sensor S1 may be received via adapted AD1. The collected sets of heart rate data may be stored to a memory MEM.

The collected sets of heart rate data may be processed with aid of a processor UP in accordance to the executable instructions APP.

The calculated PTT or any other data may be stored to the memory MEM and/or outputted via user interface Ul.

Figure 3B describes an alternative embodiment of the apparatus.

The first set of heart rate data HR1 may be collected by the first sensor S1. The first sensor S1 may include a first processor uP1 with first executable instructions and a first memory MEM1. The first processor uP1 may be used to process the first set of heart rate data HR1 based on the first instructions and store the result — in the first memory MEM1. The second set of heart rate data HR2 may be collected by the second sensor S2. The second sensor S2 may include a second processor uP2 with second executable instructions and a second memory MEM2. The second processor uP2 may be used to process the second set of heart rate data HR2 based on the second instructions and store the result in the second memory MEMZ2. A third processor uP3 may be used to collect the first set of heart rate data HR1 and the second set of heart rate data HR2 from the first memory MEM1 and the second memory MEM2. Alternatively, the third processor uP3 may be used to collect the first set of heart rate data HR1 and the second set of heart rate data HR2, or part of the first set of heart rate data HR1 and part of the second set of heart rate data N HR2 directly from the first processor uP1 and/or the second processor uP2. N The third processor uP3 may collect the first set of heart rate data HR1 and S the second set of heart rate data HR2, or part of the first set of heart rate data & HR1 and part of the second set of heart rate data HR2, directly from the first z 30 sensor S1 and/or the second sensor S2. The collected sets of heart rate data so may be processed with aid of the third processor uP3 in accordance to the N executable instructions APP.

The calculated PTT or any other data may be S stored on memory MEMS, provided for use for the processors uP1 and uP2 and/or outputted via user interface UI.

According to an exemplary embodiment, a wrist device for PPG may collect PPG based IBI2 data, and store it to a cloud. Chest strap based sensor may collect ECG data and detect RRI as IBI1, send it to a mobile phone, and the mobile phone may be configure to store the IBI1 data to the cloud. All PTT analysis may be done in the cloud. Therefore, there needs to be zero collaboration between the sensors in real time. They are just worn by the same user simultaneously. Even sensor brands can be completely independent. Timeseries data measured by any of the sensors may be stored on the sensor itself, the apparatus or the data may be transmitted into a remote storage device. Transmission of data may be wired or wireless. Wireless data transmission may be Bluetooth, wireless local area network, 3G/4G/5G mobile network or any other radio wavelength transmission. Data transmission may be implemented optically.

Timeseries data may be pre-processed in any or some of the sensors, in the apparatus, or in the storage device. Furthermore, all acguired data may be processed with aid of a processor YP in any or some of the sensors, in the apparatus, or in the storage device. These options are exemplarily, but not exclusively, described in figures 3A and 3B. A program product for determining a pulse transit time PTT may include executable instructions, software, program or code, for example. The program product may include instructions for identification and compensation of missing values. The program product for determining a pulse transit time may include N instructions for template matching. The program product for determining a N pulse transit time may include calculations for identification and removal of S artefacts. The program product for determining a pulse transit time may include & calculations for pulse transit time. The program product APP may comprise I 30 several applications, blocks or parts APP1, APP2, APP3, APP4 comprising so instructions for implementing the PTT calculation based on two simultaneously N measured sets of heart rate data.

S N Figures 4A and 4B illustrate, by way of an example, two different embodiments — for providing pulse transmit time PTT. The embodiments may illustrate method steps, separate program blocks, executable instructions or parts of algorithm for providing PTT. Figures 4A and 4B both illustrate a user as a source of information. Sets of heart rate data of a user may be measured. Two simultaneous measurements may provide IBI1 and IBI2, as input for PTT apparatus or method. IBI1 may be a measurement of electrical or mechanical activity of the heart. IBI2 may be a measurement of mechanical activity of the heart. The IBI1 and IBI2 calculations may be included in the PTT apparatus (Fig. 4A). An IBI1 apparatus may consist of IBI1 calculations, and IBI2 apparatus may consist of IBI2 calculations separate from the PTT apparatus (Fig. 4B).

IBI1 input of Figures 4A and 4B illustrate the primary R-to-R interval (RRI) seguence or pulse period P-to-P interval (PPI) seguence, which may be measured with an electrocardiography, ECG, device Or photoplethysmography, PPG, device, respectively. The RR or PP intervals may be expressed in milliseconds or seconds. RR and PP may be used interchangeably to describe beat-to-beat variability of the heart beat. Time of IBI1, tIBI1 of Figures 4A and 4B, may be a time vector for the IBI1 seguence. The time vector may correspond to the timestamps of the R-peaks in the ECG signal or P-peaks in the PPG signal. The time vector may be calculated as the cumulative sum of the IBI1 sequence. In this case, correction of missing RR or PP intervals may be implemented in order to avoid missing RR or PP intervals in the IBI1 sequence. Some of the RR or PP intervals may not be properly detected, for example due to poor signal guality. In such case, correction may include insertion of “filler” RR or PP intervals into correct N positions in the IBI1 sequence such that the cumulative sum of the IBI1 N seguence corresponds to the true elapsed time. The filler RR or PP intervals S may be selected such that they are unphysiologically long in duration ensuring & that they may be classified as artefacts by the selected artefact detection I 30 algorithm. For example, if there has been a 12-second break between two so consecutive detected R or P peaks, the missing time may be filled with, e.g., N any of the following options: S i. One 12000 ms N ii. 5000, 5000, 2000 ms iii. 5000, 3500, 3500 ms

Alternatively, the time vector may consist of the R-peak or P-peak timestamps. Correction of time vector may be implemented in order to avoid missing timestamps. Missing timestamps means that there are no time stamps for all R-peaks or P-peaks. In order to implement the correction, the IBI1 sequence may be reconstructed via taking a first-order difference of the elements in the timestamp vector. This may be followed by the optional splitting of the overly long intervals. Splitting may be implemented similarly as in the example on the splitting of the 12-second interval above (i, ii, iii). An artefact indicator sequence for the IBI1 sequence ailBI1 is presented in Figures 4A and 4B. The ailBl1 sequence may comprise Boolean values, one value per each RR or PP interval in the IBI1 sequence. The indicator value may be set to TRUE (/FALSE) if the corresponding RR or PP interval in the IBI1 sequence is marked as an artefact (/non-artefact), correspondingly, by a selected artefact detection algorithm. The selection of the artefact detection algorithm is trivial to a person skilled in the art. IBI2 input of Figures 4A and 4B illustrate the secondary inter-beat interval, IBI2, sequence, which may be measured at a peripheral location (e.g., fingertip, earlobe or wrist) using a method that detects the mechanical pulsation of the arteries. The inter-beat intervals may be expressed in milliseconds or seconds. A time vector for the IBI2 sequence tIBI2 corresponds to the timestamps of the pressure pulses detected by the sensor. The same considerations mentioned above for the time vector of IBI1 apply for the tIBI2, as well.

S N An artefact indicator seguence for the IBI2 seguence ailBI2. The same S considerations mentioned above in the context of artefact indicator of IBI1 & sequence ailBI1 apply for the ailBI2, as well. I 30 so Sub-sequences sublBl1 are identified. Identification is based on IBI1 and N ailB11. IBI1 and ailBI1 may be illustrated as vectors of certain length and the S IBI1 sequence contains all RR or PP intervals from the entire measurement or from the selected window (typically minutes or hours in duration). The sub- sequences sublBI1 are contiguous subsets of the IBI1 sequence containing only non-artefact RR or PP intervals. The length of each subsequence is much shorter than the length of the whole IBI1 sequence (typically 10-100 RR or PP intervals). The length may be predetermined, for example between determined minimum and maximum lengths; or with aid of range of minimum values and range of maximum values. Lengths and starting points of the sub-sequences may be identified. Consecutive sub-sequences may be partially overlapping. The sub-sequences may consist only valid RR or PP intervals, i.e. non- artefact. Similarly, sub-seguences sublBI2 may be identified. For each sublBI1, as identified, the best matching sublBI2 is found. This may be done with aid of an error function such that value of the chosen error function is minimized. Sub-seguences may be matched via different methods, like a method for finding the best-matching subseguence pairs, a linear-time search algorithm or a method based on dynamic programming. After synchronization shift is calculated for each valid (non-artefact) sub-seguence in IBI1, the matched inter-beat intervals in the two sequences correspond to the same heart beat intervals. In case some matched inter-beat intervals do not correspond to the same heart beat intervals, such outliers may be identified and excluded from further analysis. The outliers may be due to poor signal quality in either of IBI1 or IBI2 sequences.

Outlier pairs, where IBI1 and IBI2 do not correspond to the same heart beat intervals, may be identified. Error function values may be utilized for excluding outliers. Pairs having high error function values may be excluded directly. Further, outlier pairs to be excluded may be identified with aid of shift values, forexample compared to a mean shift value.

S N With the matched IBI1-IBI2 sequences, from which outliers are excluded, PTT S is calculated. PTT may be outputted to the user, as shown in Figures 4A and N 4B. I 30 so According to an embodiment the pulse transit time, PTT algorithm may N comprise the following four phases:

S N 1. Using IBI1 and ailBI1 (both of which are vectors of length len ibi1), find the starting indices (i ibi1 clean, a vector of length len ibi1 clean) and lengths (lengths ibi1 clean, a vector of length len ibi1 clean) of all subsequences that consist of only valid (i.e., non-artefact) RR or PP intervals and whose lengths are in the range [MIN LEN SUB, MAX LEN SUB]. a. If the number of consecutive valid (non-artefact) RR or PP intervals starting at the index i (i.e., at the element IBI1[i]) is larger than MAX LEN SUB, the value of MAX LEN SUB is set to the corresponding position in the vector lengths ibi1 clean. If, on the other hand, the number of consecutive valid RR or PP intervals starting at the index i is smaller than MIN LEN SUB, the index i is not included in the sequence i ibi1 clean, and the corresponding length (which is now smaller than MIN LEN SUB) is not included in the vector lengths ibi1 clean. b. The optimal values for the constants MIN LEN SUB and MAX LEN SUB often depend on the IBI1 data quality, i.e., on the artefact percentage and on the distribution of artefacts within the IBI1 sequence. Good values are often found in the ranges MIN LEN SUB = 3-20 and MAX LEN SUB = 10-100 (while ensuring that MAX LEN SUB never falls below MIN LEN SUB). c. NB: the consecutive subseguences may be partially overlapping if the artefact percentage of the IBI1 sequence is very low. E.g., if the artefact percentage equals 0 %, and len ibi1 = 1000 and MIN LEN SUB = 10, and MAX LEN SUB = 20; the index vector i ibi1 clean contains the values [0,1,2,...,980,981,982,...,990], and the length vector lengths ibi1 clean contains the corresponding values [20,20,20,...,20,19,18,...,10]. Both vectors are now of length len ibi1 clean = 991. NB: a zero-based indexing scheme is assumed N here. & S 2. Similarly, using IBI2 and ailBI2 (both of which are vectors of length len ibi2), & find the starting indices (i ibi2 clean, a vector of length len ibi2 clean) and I 30 lengths (lengths ibi2 clean, a vector of length len ibi2 clean) of all so contiguous subsequences that consist of only valid (i.e., non-artefact) inter- N beat intervals and whose lengths are in the range [MIN LEN SUB, S MAX LEN SUB].

N

3. For each clean IBI1 subsequence starting at index i ibi1 clean[i] (where the index i is in range [0, len ibi1 clean - 1]), find the best matching clean subsequence in IBI2 in terms of a selected error function.

If the clean subsequence in IBI2 starting at location i ibi2 clean[j] (where the index j is in range [O, len ibi2 clean - 1]) is compared with the IBI1 subsequence starting at index i ibi1 cleanjfi], the value of the selected error function should be calculated using len ij = MIN(lengths ibi1 clean[i], lengths ibi2 clean[j]) elements from each IBI1/IBI2 sequence starting at the respective locations.

The best matching subseguence in IBI2 (i.e., the starting index j) is selected such that the value of the chosen error function is minimized. a.

Good candidates for the error function are, e.g., the following: i.

Root mean squared error (RMSE) ii.

Mean absolute error (MAE) iii.

Mean absolute percentage error (MAPE) b.

NB: it may be wise to modify the error function such that short subseguences (i.e., low values for len ij) are penalized since the probability of finding a spurious well-matching subseguence gets higher as the len ij gets smaller.

Such penalization may be achieved, e.g., via adding a non-negative penalty term to the chosen error function.

The penalty term may be, e.g., of the form PENALTY_CONST / (len_ijj*EXP), where PENALTY CONST and EXP are user-defined constants (scaling factor and exponent). Good candidates for these constants may be found in the ranges PENALTY CONST = 1.0 - 1000.0 and EXP = 0.1 - 5.0. c.

Now that the index pair (i, j) has been fixed (i.e., the subsequence of length len ij starting at IBI2[j] yields smallest error function value with the reference subsequence starting at IBI1[i]), the optimal temporal shift for N the secondary IBI2 sequence at the time point time ibi2[j] is shift ij = N time ibi1[i] - time ibi2[j]. If the offset shift ij is added to each element of S the time vector time ibi2, it is expected that the IBI1 and IBI2 sequences & are well-synchronized (i.e., the corresponding inter-beat interval values I 30 are nicely overlapping when the two sequences IBI1 and IBI2 are plotted so against their respective time vectors in the same figure) at (or near) the N time instant time_ibi1[i]. S d.

There are various possibilities for the implementation of the subsequence-matching (aka template matching) phase of the PTT algorithm.

The most straightforward method for finding the best- matching subsequence pairs is the exhaustive search method, where all elements in i ibi2 clean are scanned for each element in i ibi1 clean. However, such is also computationally the most demanding with guadratic time complexity. If, on the other hand, there is some prior information available on the approximate range in which the shift ij values reside, one may limit the scanning range such that the timestamps of the compared subseguences are never too far outside this range given as prior information. This yields a linear-time search algorithm since constant amount of work is done for each element in i ibi1 clean (only a subset of i ibi2 clean of constant size is scanned). Alternatively, a method based on dynamic programming may be utilized, where the sum of the error function values (one value per (i, j) pair, where i is in i ibi1 clean and j is in i ibi2 clean) is minimized with the additional reguirement that the seguence of indices j is monotonically increasing. e. When the optimal synchronization shift is calculated for each clean subseguence in IBI1, vast majority of the calculated synchronization shifts correspond to the “true” temporal shift between the two signals (i.e., the matched inter-beat intervals in the two seguences correspond to the same heart beat intervals). There may be occasional outliers where the matched inter-beat intervals do not correspond to the same heart beat intervals, e.g., due to occasional poor signal guality in either one (or both) of the IBI1/IBI2 sequences. The outliers must be identified and excluded from further analysis. This will be done in the next step.

4. The outlier pairs (where the RR or PP intervals in the subseguence starting at IBI1[i] do not correspond to the same underlying heart beats as the inter- N beat intervals in the subsequence starting at IBI2[j])) may be identified, e.g., N using the error function values calculated for the pairs: pairs with S exceptionally high error function values may be excluded directly (an error & function dependent threshold value must be heuristically devised for this I 30 procedure). In addition, there may be outlier pairs with very small error so function values (a well-matching sequence was found by chance from an N erroneous location in the IBI2 sequence). These outliers may be identified S based on the shift values (shift ij): the shift value is classified as an outlier if it deviates greatly from a mean shift value which may be calculated, e.g., as the weighted average of shift_ij values using the transformed error function values (e.g., using a downward-sloping exponential function) as weights. A temporally smooth function (e.g., a spline) may finally be interpolated through the retained time points: time_ibi1[i] on the ordinate and shift_ij on the abscissa. This smooth interpolated function then displays the temporal variation in the pulse transit time.

a. NB: The simple data selection method described above may be too simple for such cases where the artefact percentage of either of the IBI1 or IBI2 sequence (or both) is very high: the sheer number of outliers may distort the accuracy of the weighted average mentioned above. In such cases, the weighted average of the shift value may be calculated using the iteratively re-weighted scheme, where — on each iteration — the shift values are additionally down-weighted based on their distance from the weighted mean calculated on the previous iteration. The down-weighting may be carried out, e.g., using a Gaussian kernel whose kernel width (the Gaussian sigma) is decreased exponentially (multiplied with a constant in the open range (0.0, 1.0)) on each iteration while preventing it from falling below a predetermined lower limit. Natural stopping conditions for the iterative algorithm are, e.g., when the estimated value has changed by less than a pre-defined limit since the previous iteration, or when the number of iterations exceeds a pre-defined limit.

b. Sometimes, there might be an error (a constant offset) in the clock frequency either one (or both) of the two devices (an ECG device or PPG device and the peripheral device which may be based on PPG). The constant offset in the clock frequency linearly stretches (if the clock frequency is too high) or contracts (if the clock frequency is too low) the time vector of one IBI1/IBI2 sequence in relation to the time axis of the N other IBI1/IBI2 sequence. If the error in the clock frequency remains N constant, the optimal temporal shift (shift ij at the time point time rrili]) S changes linearly over time. The linear trend is increasing (decreasing) if & the clock freguency of the secondary peripheral device is too low (too Ek 30 high) or if the clock frequency of the primary ECG or PPG device is too so high (too low). Usually this linear effect due to the constant error in clock N freguency is much larger in magnitude than the shift variations caused by S variations in pulse transit time. The PTT effect is only seen as a weak ripple on top of the linear trend. Thus, the linear trend must be removed so that the PTT information may be studied. First, the linear temporal trend in the optimal shift (i.e., the offset and slope parameters of the linear function approximating the linear trend) must be estimated from the data. i.

The iteratively re-weighted scheme described above in the context of the mean shift value may be used also in simple linear least squares regression (IRWLS, iteratively re-weighted least squares regression). First, the slope and offset are estimated using all data (with perhaps the most obvious outliers excluded) using weighted linear least squares regression (using the transformed error function values as weights, as usual). Then, on subsequent iterations, the parameters are re-estimated using re-weighted data points.

The re-weighting is based on their distances from the estimated linear line (using the Gaussian kernel with exponentially decreasing kernel width, as mentioned above). The iteration is stopped whenever the estimated parameters no longer change or when the maximum number of iterations is reached. i.

Once the linear trend — i.e., the offset (offset trend) and slope (slope trend) parameters of the linear function — has been estimated, its effect may be eliminated from the shift ij values plot via subtraction: the corrected value at time point time_ibi1[i] will be (shift ij - offset trend - slope trend*time ibi1[i]). Alternatively, the trend may be eliminated via dividing each element of the IBI2 sequence and each value of the time ibi2 vector by the value (1 - slope trend), after which the entire analysis may be repeated using the preprocessed (i.e., stretched or contracted) IBI1 and IBI2 sequences.

This re-scaling operation stretches (contracts) the time ibi2 vector if the value of N slope trend is negative (positive), i.e., if the clock freguency of the N corresponding sensor is too high (low). S iii.

If the error in the clock frequency does not remain constant in time, a & non-linear function (such as higher order polynomial) may be used to I 30 estimate the baseline caused by the drift.

Similar IRWLS algorithms so are available for non-linear functions, as well.

N iv.

It should be noted that the two sensors (the primary ECG or PPG S sensor and the peripheral sensor) need not be synchronized for PTT analysis since the synchronization is done by the current algorithm.

In addition, slowly varying or constant errors in the clock frequencies are permitted, as well, as the trend-correction method will eliminate the disturbance caused by the error. The proposed algorithm may naturally be used in offline analysis. In this case, the IBI1 and IBI2 sequences are first collected after which the analysis is carried out. The resulting PTT time series may then be visualized and the blood pressure variability BPV trend may be extracted from it. However, real-time analysis is also possible with the proposed algorithm. In this case, either the entire IBI1 and IBI2 sequences from the beginning of the measurement or only the most recent data from a predetermined time window (say, 30 minutes to 8 hours) are kept in memory, and the analysis is performed using the data stored in memory. Figure 5 illustrates the detection of missing or extra pulses in the heart rate — timeseries. In the top panel a simplified pulse wave or primary ECG/PPG signal, where each peak corresponds to a heartbeat, is presented. The pulse wave signal is an evenly spaced time series. Usually the sampling rate is close to 1000 Hz to permit 1 ms temporal resolution. Inter-beat intervals (RR intervals in the case of ECG) are marked with arrows above the pulse wave. The duration of each inter-beat interval is given in milliseconds above each arrow. The cumulative sum of the inter-beat intervals (i.e., the time vector of the original inter-beat interval seguence) is given near the bottom of the panel. Two missing heartbeats are drawn with dashed line (three inter-beat intervals are replaced with a single long 2684 ms interval) and one extra beat is located attime 29389 ms: one 1054 ms inter-beat interval is erroneously split into two N short intervals (579 and 475 ms).

N S The middle panel of the Figure 5 presents the corresponding inter-beat interval & seguence plotted against the cumulative sum (which is used as time vector). I 30 The time vector used in this middle panel is unevenly spaced. Three inter-beat so intervals in this sequence are classified as artefacts by the artefact detection N algorithm. The artefacts are marked with filled circles. The 2684 ms inter-beat S interval is classified as an artefact since it is unphysiologically long in duration. The two short inter-beat intervals (579 and 475 ms) are also classified as artefacts since they are much shorter than the surrounding inter-beat intervals.

In the bottom panel of Figure 5 the original inter-beat interval sequence (middle figure) is re-sampled into an evenly spaced time series with the sampling interval of 100 ms. A zeroth order spline is used here as the interpolation method. In other words, the last measured non-artefact inter-beat interval is used as the interpolated value at each temporal sampling point. Other interpolation methods (such as linear or polynomial splines) are also possible. It is seen that the inter-beat intervals classified as artefacts were not included in the interpolation.

In steps 1-3 of the proposed algorithm, the IBI1 and IBI2 sequences may also be re-sampled into an evenly-spaced time series via selecting a sampling interval from the range of, e.g., 50 - 500 ms. The re-sampling may be carried out with any interpolation method, e.g., using temporally nearest or previous non-artefact IBI2/IBI1 value at each temporal sampling point. Alternatively, — locally linear or polynomial splines may be used to determine the interpolated value at each sampling point. The IBI1s/IBI2s classified as artefacts should not be used in the interpolation step. The interpolated values at each sampling point should be determined only using IBI1s/IBI2s that were classified as non- artefacts by the selected artefact detection algorithm. Illustration of artefact detection, interpolation of missing values and removal of extra IBI1s/IBI2s is presented in Figure 5. The use of evenly-spaced IBI1/IBI2 time series permits the use of digital signal processing methods designed for evenly spaced time series. E.g., the template matching step (where the optimal shift of each subseguence is determined) may be carried out via convolution for which computationally efficient (sub-guadratic) algorithms exist.

S N The blood pressure is known to show spontaneous, detectable changes over S very short time periods. The changes may be measured and described as & blood pressure variability, BPV. When the blood pressure decreases, the pulse x 30 transit time, PTT increases, and vice versa.

LO N The blood pressure variability has been found to be an indicator of significant S cardiovascular events and some initial investigations have shown that measurement of the BPV may be a predictor of an endurance performance as aresult of intensive training. In addition, BPV may be used as an indicator of autonomic regulation and/or to detect abnormalities in the cardiovascular system. By measuring the changes in the pulse transit time, it is possible to provide information on blood pressure variability. In the previous description aspects of the invention are illustrated with exemplary embodiments, figures and examples. Some blocks or parts, like devices or sensors, may be changed between each other or to another kind, or even left out without departing from the scope of the invention as determined in the following claims.

O QA O N

K <Q

O N

I a a

LO N NN LO O QA O N

Claims

Claims:

1. Method for determining a change in pulse transit time, comprising - detecting first activity of a heart rate of a user, - detecting second activity of the heart rate of the user, - determining sequences of the first activity from the first activity of the heart rate as function of time, - determining sequences of the second activity from the second activity of the heart rate as a function of time, - identifying sequences of the first activity and sequences of the second activity corresponding to the same heart beat by - determining sub-sequences of the first activity of the heart rate as contiguous subsets of sequences of the first activity of the heart rate, and sub-sequences of the second activity of the heart rate as contiguous subsets of sequences of the second activity of the heart rate, - identifying length and starting point of individual sub- sequence of the first activity of the heart rate, and length and starting point of individual sub-sequence of the second activity of the heart rate, - identifying best match for individual sub-sequence of the first activity of the heart rate in relation to individual sub-sequence of the second activity of the heart rate, and - calculating synchronization shift or offset of the sub- sequence of the first activity of the heart rate at matching N point of time, N - calculating the average mutual offset of the sequences of the first S activity and seguences of the second activity, = - calculating the change in the offsets of the sequences of the first I 30 activity and sequences of the second activity, and so - determining the changes in the pulse transit time, PTT as a N difference between timely matched seguences of the first activity S and seguences of the second activity.

N

2. Method according to the claim 1, comprising detecting the first activity of the heart a. by an electrical sensor or an electrocardiogram or b. by a mechanical sensor, a photoplethysmogram, a ballistocardiogram or a pressure sensor.

3. Method according to any of the claims 1-2, comprising detecting the second activity of the heart via a mechanical sensor, a photoplethysmogram, a ballistocardiogram or a pressure sensor.

4. Method according to any of the claims 1-3, comprising determining the sequences of the electrical activity via cycle lengths or intervals of successive R-peaks of the electrocardiogram.

5. Method according to any of the claims 1-4, comprising detecting the sequences of the mechanical activity via at least one of: pulsation of a blood pressure, volume in peripheral area, successive timestamps of the peaks of mechanical pulsation of the arteries, and inter-beat intervals.

6. Method according to any of the claims 1-5, wherein determining the sequences of the first activity as a function of time and determining the sequences of the second activity as a function of time includes at least one of: - providing time axis for the sequences of the first activity and for the sequences of the second activity, - creating time axis by a cumulative sum of the sequences of the first N activity and by a cumulative sum of the seguences of the second N activity, S - calculating time vectors for the first activity and for the second = activity, I 30 - calculating time vectors for the sequences of the first activity and for so the sequences of the second activity, optionally by cumulative sums, N - correcting for missing values in the first activity of the heart rate S based on cumulative sum of the sequence of the first activity or first- order difference or any other high-pass filtering method of the detected elements in the first activity of a heart rate,

- correcting for missing values in the second activity of the heart rate based on cumulative sum of the sequence of the second activity or first-order difference of the detected elements in the second activity of the heart rate - re-sampling the first and second sequences of the activity of the heart into evenly-spaced artefact-corrected time series for which evenly-spaced time vectors may be constructed

7. Method according to any of the claims 1-6, comprising matching the sequences of the first activity and the sequences of the second activity with each other timely via linear or non-linear fitting, optionally taking into account the device internal clock offset between a device arranged to measure the first activity and a device arranged to measure the second activity.

8. Method according to any of the claims 1-7, comprising detecting and removing artefacts from the sequences of the first activity, and from the sequences of the second activity.

9. Method according to any of the claims 1-8, comprising identifying starting locations of sub-sequences of the first activity as a function of time, and of sub-sequences of the second activity as a function of time, optionally the sub-sequences including a predetermined length or a predetermined number of elements.

N N

10. Method according to the claim 1, comprising identifying a matching sub- S seguence of the second activity for each sub-seguence of the first = activity via error metrics, optionally via minimum mean absolute x 30 percentage error.

LO N

11. Method according to the claims 1 and 10, comprising shifting, stretching S or contracting one of the following:

O

N a. one or more sub-seguences of the second activity as a function of time based on the sub-seguences of the first activity; and b. one or more sub-sequences of the first activity as a function of time based on the sub-sequences of the second activity.

12. An apparatus for determining a change in pulse transit time, comprising a. a memory for storing arrangements and a processor, b. afirst sensor configured to detect electrical activity of a heart rate of a user or mechanical activity of the heart of the user, c. a second sensor configured to detect mechanical activity of the heart rate of the user, d. an arrangement configured to determine sequences of the first activity from the first activity of the heart rate as function of time, e. an arrangement configured to determine sequences of the second activity from the second activity of the heart rate as a function of time, f. an arrangement configured to identifying sequences of the first activity and sequences of the second activity corresponding to the same heart beat with - an arrangement configured to determine sub-sequences of the first activity of the heart rate as contiguous subsets of sequences of the first activity of the heart rate, and sub- sequences of the second activity of the heart rate as contiguous subsets of sequences of the second activity of the heart rate, - an arrangement configured to identify length and starting point of individual sub-sequence of the first activity of the N heart rate, and length and starting point of individual sub- N seguence of the second activity of the heart rate, S - an arrangement configured to identify best match for = individual sub-seguence of the first activity of the heart I 30 rate in relation to individual sub-sequence of the second so activity of the heart rate, N - an arrangement configured to calculate synchronization S shift or offset of the sub-sequence of the first activity of the heart rate at matching point of time,

g. an arrangement configured to calculate the average mutual offset of the sequences of the first activity and sequences of the second activity, h. an arrangement configured to calculate the change in the offsets of the sequences of the first activity and sequences of the second activity, and i. an arrangement configured to determine the changes in the pulse transit time, PTT as a difference between timely matched seguences of the first activity and seguences of the second activity.

13. An apparatus according to any of claims 1-11, comprising a memory, a processor and means for carrying out a method according to any of the claims 1-11.

14. A computer program product for determining a change in pulse transit time comprising instructions, which when executed by a processor, carry out at least the following steps

1. retrieving detected and/or stored first activity of a heart rate and second activity of a heart rate,

2. processing detected and/or stored first activity of a heart rate and second activity of a heart rate by

3. determining seguences of the first activity from the first activity of the heart rate as function of time, N 4. determining seguences of the second activity from the second N activity of the heart rate as a function of time, S 5. identifying seguences of the first activity and seguences of the = second activity corresponding to the same heart beat by z 30 - determining sub-seguences of the first activity of the heart so rate as contiguous subsets of sequences of the first activity of N the heart rate, and sub-seguences of the second activity of S the heart rate as contiguous subsets of sequences of the second activity of the heart rate, - identifying length and starting point of individual sub- sequence of the first activity of the heart rate, and length and starting point of individual sub-sequence of the second activity of the heart rate, - identifying best match for individual sub-sequence of the first activity of the heart rate in relation to individual sub-sequence > of the second activity of the heart rate, and - calculating synchronization shift or offset of the sub-seguence of the first activity of the heart rate at matching point of time,

6. calculating the average mutual offset of the seguences of the first activity and sequences of the second activity,

7. calculating the change in the offsets of the seguences of the first activity and seguences of the second activity, and

8. determining the changes in the pulse transit time, PTT as a difference between timely matched seguences of the first activity and seguences of the second activity.

N

O

N

O <Q

NN

I a a

LO

N

NN

LO

O

QA

O

N