CN114469127A - Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding - Google Patents
Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding Download PDFInfo
- Publication number
- CN114469127A CN114469127A CN202210308702.0A CN202210308702A CN114469127A CN 114469127 A CN114469127 A CN 114469127A CN 202210308702 A CN202210308702 A CN 202210308702A CN 114469127 A CN114469127 A CN 114469127A
- Authority
- CN
- China
- Prior art keywords
- differential
- heartbeat
- neural network
- template
- heart beat
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
- 238000012545 processing Methods 0.000 title claims abstract description 27
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 21
- 238000013528 artificial neural network Methods 0.000 claims abstract description 63
- 230000000747 cardiac effect Effects 0.000 claims abstract description 32
- 230000004913 activation Effects 0.000 claims abstract description 30
- 230000002159 abnormal effect Effects 0.000 claims abstract description 21
- 238000007781 pre-processing Methods 0.000 claims abstract description 19
- 238000003062 neural network model Methods 0.000 claims description 55
- 238000012549 training Methods 0.000 claims description 24
- 238000012795 verification Methods 0.000 claims description 12
- 238000005070 sampling Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000013145 classification model Methods 0.000 claims description 6
- 230000010247 heart contraction Effects 0.000 claims description 6
- 230000002618 waking effect Effects 0.000 claims description 6
- 230000003044 adaptive effect Effects 0.000 claims description 4
- 230000001186 cumulative effect Effects 0.000 claims description 2
- 238000000034 method Methods 0.000 abstract description 8
- 230000008569 process Effects 0.000 abstract description 3
- 238000002565 electrocardiography Methods 0.000 description 19
- 238000012544 monitoring process Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 206010003119 arrhythmia Diseases 0.000 description 4
- 230000006793 arrhythmia Effects 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 230000005189 cardiac health Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 208000024172 Cardiovascular disease Diseases 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000012806 monitoring device Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000005182 global health Effects 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000033764 rhythmic process Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
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]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses an artificial intelligence processing circuit for electrocardiosignals based on cardiac beat differential coding, and belongs to the technical field of electrocardiosignal processing. The scheme is as follows: the preprocessing module is connected with the differential coding module through the heartbeat memory, the differential coding module is sequentially connected with the activation module and the neural network classifier, the differential network memory and the non-differential network memory are connected with the neural network classifier and connected with the heartbeat template memory through the template updating module, and the heartbeat template memory is further connected with the differential coding module. The invention utilizes the waveform characteristics and the self-adaptive threshold value of the differential heartbeat to realize the dynamic awakening of the normal abnormal heartbeat, thereby reducing the starting times of the artificial intelligent classifier and finally reducing the power consumption of the whole system. According to the invention, the patient heart beat template is generated by using the historical classification result of the artificial intelligent classifier, the generalization accuracy of model processing is improved, the patient heart beat does not need to be additionally labeled in the whole process, and the self-specificity of the patient can be learned in real time.
Description
Technical Field
The invention belongs to the technical field of electrocardiosignal processing, and particularly relates to an artificial intelligence electrocardiosignal processing circuit based on cardiac beat differential coding.
Background
Cardiovascular disease (CVD) is the leading cause of death worldwide according to the 2019 global health assessment report issued by the World Health Organization (WHO). For accurate detection of CVD patients, Electrocardiography (ECG) plus manual arrhythmia analysis by a specialist are generally used medically, but such solutions are difficult to use in the field of home medical equipment due to the cumbersome equipment and high labor costs. In recent years, the household medical field is more and more popular with the coming of the intelligent times, and in the aspect of ECG heart health monitoring, daily electrocardiographic monitoring equipment with intelligent analysis capability receives wide attention: 1) compared with hospital examination, the daily ECG monitoring can find arrhythmia problems in a long-term and timely manner; 2) because early abnormal heart activity is often sporadic in character, hospital short-term ECG monitoring can hardly detect such problems, and eventually the patient may miss the gold treatment period.
To meet market expectations, an excellent daily cardiac health monitoring device needs to meet the requirements of real-time, intelligence, wearability, long endurance, and high accuracy simultaneously: real-time and intelligent guarantee in time follow the discovery arrhythmia phenomenon in the ECG waveform, and wearable design reduces the influence to user's daily life, and long duration avoids too frequent monitoring interrupt because of changing the battery or charging leads to, and final high accuracy is the basic requirement to this type of equipment. In general, the five requirements are often restricted, and it is difficult to simultaneously achieve the requirements by directly designing based on the existing hardware, so that an Application Specific Integrated Circuit (ASIC) chip designed by software and hardware in cooperation becomes a current research hotspot, and many advanced ASIC chips for physiological signal identification in the world seek lower power consumption level to meet the long-term endurance requirement.
Furthermore, applying prior art achievements to actual daily cardiac health monitoring devices also requires addressing the bottleneck of generalized accuracy: in the field of ECG monitoring, the intrinsic difference of heart cycle activity of each person is caused by the difference of physiological structures between patients, so that the ECG waveform for recording the heart cycle activity has specificity among patients, and in addition, the self-specificity of the patient for the ECG recording of the patient also has the self-specificity of the patient which changes along with time due to the position difference or age change of the heart electrodes. However, for arrhythmia identification, the artificial intelligence classifier can only contact the ECG records of patients in the database for a period of time in the learning stage, so that the conventional artificial intelligence classifier usually shows extremely high identification accuracy on the database, but the problem that the generalization accuracy is greatly reduced for patients outside the database possibly occurs. Existing solutions such as online learning also have inherent problems: 1) they all require the acquisition of a portion of the ECG recording of the actual user (patient) and labeling the ECG waveform with a classification label for each heartbeat by a specialist; 2) the inability to learn the latest heartbeat characteristics (i.e., ECG characteristics) of a patient in real time may lead to decreased accuracy in the user's daily use as the user is re-worn.
Disclosure of Invention
The invention provides an artificial intelligence processing circuit of electrocardiosignals based on cardiac beat differential coding, which can be used for improving the accuracy of classification and identification of the electrocardiosignals.
The invention provides an artificial intelligence processing circuit of electrocardiosignals based on heart beat differential coding, which comprises: the device comprises a preprocessing module, a heartbeat memory, a heartbeat template memory, a differential coding module, an activation module, a neural network classifier, a differential network memory, a non-differential network memory and a template updating module, wherein the preprocessing module is connected with the differential coding module through the heartbeat memory, the differential coding module is sequentially connected with the activation module and the neural network classifier, the differential network memory and the non-differential network memory are connected with the neural network classifier, the neural network classifier is connected with the heartbeat template memory through the template updating module, and the heartbeat template memory is connected with the differential coding module;
the preprocessing module is used for preprocessing the input electrocardiosignals, detecting the heartbeat signals, extracting single heartbeat signals and storing the extracted heartbeat signals to the heartbeat memory;
the non-differential network memory is used for storing a pre-trained non-differential neural network model M1, and the differential network memory is used for storing a pre-trained differential neural network model M2; the non-differential neural network model M1 and the differential neural network model M2 are used for carrying out two classification recognition processing of normal heartbeat and abnormal heartbeat on the input heartbeat signal;
the differential encoder comprises two working modes, which are respectively:
in the first working mode, the differential encoder reads a heartbeat signal from the heartbeat memory as a first heartbeat signal and sends the first heartbeat signal to the neural network classifier, and when the neural network classifier detects that the currently received heartbeat signal is the first heartbeat signal, the neural network classifier reads a non-differential neural network model M1 from the non-differential network memory as a current classification model for identification and classification to obtain a first classification result and sends the first classification result to the template updating module; the template updating module takes the average value of normal heartbeats in the first classification results as an initial heart beat template and stores the initial heart beat template into a heart beat template memory based on the obtained plurality of first classification results;
in the second working mode, the differential encoder reads the heart beat signal currently obtained by the preprocessing module from the heart beat memory in real time, obtains a differential heart beat signal based on the difference value between the current heart beat signal and the current heart beat template, and sends the differential heart beat signal to the activation module; the activation module is used for calculating the variance of the differential heartbeat signal, if the variance of the current differential heartbeat signal is larger than a threshold Th, the neural network classifier is awakened, and the current differential heartbeat signal is sent to the neural network classifier; if the variance of the current differential heartbeat signal is less than or equal to the threshold Th, the classification result of the current heartbeat signal is normal heartbeat, and the classification result of the current heartbeat signal is directly output;
when the neural network classifier receives the differential heartbeat signal, reading a differential neural network model M2 from a differential network memory as a current classification model for identification and classification, obtaining a second classification result and sending the second classification result to a template updating module;
when the template updating module detects that the current second classification result is normal heartbeat, calculating a weighting updating coefficient a of the current differential heartbeat signal according to a configured calculation strategy, obtaining an updating amount according to the product of the weighting updating coefficient a and the current differential heartbeat signal, and updating the heartbeat template stored in the heartbeat template memory, wherein the updated heartbeat template is as follows: the sum of the pre-update heart beat template and the update amount.
Further, the weighting update coefficient a is: the current differential heartbeat signal output by the differential neural network model M2 is judged to be one half of the difference between the probability value of normal heartbeat and the probability value of abnormal heartbeat, and then multiplied by a preset threshold value threshold.
Further, the specific calculation mode of the activation module for calculating the variance of the differential heartbeat signal is as follows:
uniformly dividing the differential heart beat signal into M sections, wherein M is an integer greater than 1; and defining the number of sampling points contained in each section as N;
for the mth section differential heart beat signal, the first N/2 sampling points are taken to calculate the mean valueX mWherein M =1,2, …, M;
respectively calculating each sampling point and the mean value of the last N/2 sampling points of the mth section differential cardiac signalX mAnd the cumulative sum of the absolute differences is recorded asY m ；
Based on M-segment difference cardiac signalsY m The mean of (d) is taken as the approximate variance and used as the variance of the differential beat signal.
Further, the activation module further comprises: when the accumulated times of not waking up the neural network classifier exceeds a frequency threshold value, directly waking up the neural network classifier, sending a latest differential heartbeat signal to the neural network classifier, waiting for the neural network classifier to feed back a classification result, and reducing the threshold value Th according to a specified proportion if the differential heartbeat signal is abnormal heartbeat.
Further, the activation module performs adaptive threshold adjustment on the threshold Th in real time according to the classification result currently output by the neural network classifier: if the classification result is normal heartbeat, updating the threshold Th according to a formula Th = Th + betax (Y-Th) x c; if the classification result is abnormal heartbeat, updating the threshold Th according to a formula Th = gamma multiplied by Y multiplied by c;
where Y denotes a variance of the differential heartbeat signal, c denotes a preset confidence factor, and β and γ denote preset first and second coefficients, respectively.
Further, the training mode of the non-differential neural network model M1 and the differential neural network model M2 includes:
(1) dividing continuous electrocardiosignals into a plurality of heart beat waveforms based on the designated characteristic positions of the electrocardiosignals in the patient electrocardiosignal recording database, thereby obtaining a heart beat data set D1, and dividing the heart beat data set D into a training set and a verification set according to a designated proportion;
(2) training network parameters of a classification neural network model of the non-differential neural network model M1 by using a training set of a heartbeat data set D1, and selecting the classification neural network model with the optimal accuracy rate through a verification set of a heartbeat data set D1 to obtain a non-differential neural network model M1;
(3) sending the first N 'heart beating signals recorded by the electrocardiosignals of each patient in the heart beating data set D1 into a non-differential neural network model M1 for classification, counting the number of normal heart beats as a classification result, and obtaining an initial heart beating template based on the mean value of K normal heart beats when K heart beats are obtained, wherein N' and K are positive integers more than 1;
(4) sequentially traversing the cardiac beat signal of the electrocardiosignals of each patient in the cardiac beat data set D1HB in Based on the current heartbeat signalHB in Obtaining a differential heartbeat signal by the difference between the current heartbeat template and the current heartbeat templateHB diff Sequentially sending the heart beat data to a non-differential neural network model M1 for classification, and updating the heart beat template based on the currently obtained classification result;
based on all differential cardiac signalsHB diff Obtaining a heart beat differential data set D2, and dividing the heart beat differential data set D2 into a training set and a verification set according to a specified proportion;
(5) and training the network parameters of the classified neural network model of the differential neural network model M2 by using a training set of the heart beat differential data set D2, and selecting the classified neural network model with the optimal accuracy rate by using a verification set of the beat differential data set D2 to obtain the differential neural network model M2.
The technical scheme provided by the invention at least has the following beneficial effects:
in the invention, the dynamic awakening of the abnormal positive heartbeat is realized by utilizing the waveform characteristics and the self-adaptive threshold of the differential heartbeat, so that the starting times of the artificial intelligent classifier are reduced, and the power consumption of the whole system is finally reduced. The invention further discloses a neural network training method based on the heart beat differential coding, and is matched with the electrocardiosignal artificial intelligence processing circuit based on the heart beat differential coding, a patient heart beat template is generated by utilizing the historical classification result of an artificial intelligence classifier, and the generalization accuracy of a network model is improved by combining the heart beat differential processing, and the whole process does not need to additionally label the heart beat of the patient and can learn the self-specificity of the patient in real time.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a schematic structural diagram of an artificial intelligence processing circuit for electrocardiosignals based on cardiac beat differential encoding according to an embodiment of the present invention;
fig. 2 is a flowchart of the operation of an artificial intelligence processing circuit for cardiac signals based on cardiac beat differential encoding in an initialization stage according to an embodiment of the present invention;
fig. 3 is a flowchart of the operation of the artificial intelligence processing circuit for cardiac signals based on cardiac beat differential encoding in the working phase according to the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
Due to the patient (user) specificity of the ECG signal, the normal ECG waveform of each person can have considerable difference due to individual difference and age change, and therefore, the relatively limited ECG recording database causes the problem that the trained model has low generalization accuracy. In the embodiment of the invention, the heart beat of the patient is divided into two parts: a portion of the basic waveform resulting from normal cardiac activity and a portion of the waveform variation resulting from abnormal cardiac activity. In the embodiment of the invention, a template generated by normal heart beat of a user is used as a basic waveform part, and the difference value between the heart beat of the current user and the basic waveform is used as a waveform variation part, so that the artificial intelligent electrocardiosignal processing circuit based on heart beat differential coding is realized.
As a possible implementation manner, the artificial intelligence processing circuit for cardiac signals based on cardiac beat differential coding adopted in the embodiment of the present invention includes: the device comprises a preprocessing module, a heartbeat memory, a heartbeat template memory, a differential coding module, an activation module, a neural network classifier, a differential network memory, a non-differential network memory and a template updating module, wherein as shown in figure 1, the preprocessing module is connected with the differential coding module through the heartbeat memory, the differential coding module is sequentially connected with the activation module and the neural network classifier, the differential network memory and the non-differential network memory are connected with the neural network classifier, the neural network classifier is connected with the heartbeat template memory through the template updating module, and the heartbeat template memory is also connected with the differential coding module.
The preprocessing module is used for preprocessing the input electrocardiosignals, and comprises: detecting the heart beat, extracting the single heart beat signal, and storing the extracted heart beat signal in a heart beat memory. Specifically, the preprocessing module realizes heartbeat detection through a detection strategy preset on the preprocessing module, and then extracts heartbeat signals according to a certain window length. Any conventional technical means can be adopted for heartbeat detection, and the embodiment of the present invention is not particularly limited in this respect.
The non-differential network memory is used for storing the pre-trained non-differential neural network model M1, and the differential network memory is used for storing the pre-trained differential neural network model M2. The model M1 and the model M2 are used for performing two classification recognition processes of normal heartbeat and abnormal heartbeat on the input heartbeat signal, and the network structures of the model M1 and the model M2 may adopt any neural network structure commonly used in the art for classification tasks, which is not specifically limited in this embodiment of the present invention.
When the differential encoder works in a first working mode (namely an initialization stage), the differential encoder reads a heartbeat signal from a heartbeat memory as a first heartbeat signal (input data of the model M1) and sends the first heartbeat signal to the neural network classifier, and when the neural network classifier detects that the currently received heartbeat signal is the first heartbeat signal, the neural network classifier reads a non-differential neural network model M1 from a non-differential network memory as a current classification model for identification and classification, obtains a first classification result and sends the first classification result to the template updating module. The template updating module beats the normal hearts in the first classification results based on the obtained plurality of first classification resultsHB normal As an initial heart beat templateHB temp And stored in the heart beat template memory.
For example, defining the normal heart beat currently obtainedHB normal Is defined as K, then the heart beat template is initializedHB temp The calculation formula of (2) is as follows:
wherein i is a number for indicating a normal heartbeat,HB normal [i]indicating the ith normal heart beat.
When the differential encoder works in the second working mode, the differential encoder reads the heart beat signal currently obtained by the preprocessing module from the heart beat memory in real timeHB in According to the heart beat signalHB in And the current heart beat template: (HB temp ) Calculating a differential cardiac signal from the difference ofHB diff And the difference is divided into a heart beat signalHB diff To the activation module. Wherein, the heart beat signal is differentiatedHB diff Is shown in formula (2):
at the same time, the activation module will differentially beat the signalHB diff The method comprises the following steps of dividing the segments into multiple segments, defining M to represent the number of the segments after division, defining a sampling point contained in each segment as N, and representing N sampling points as: s1~sN。
Then for any segment M (M =1,2, …, M), the average can be calculated by taking the first N/2 points according to equation (3)X mAccording to formula (4) from the mean valueX mAnd calculating intermediate quantity by last N/2 pointsY mAnd finally, averaging the M sections of signals through a formula (5) to obtain an approximate variance Y for measuring the similarity of the current heart beat and the heart beat template.
The activation module compares the approximate variance Y with a preset threshold Th, and if the approximate variance Y is greater than the threshold, the neural network is activated, i.e., the differential cardiac signal is obtainedHB diff And (3) sending the data to a neural network classifier for further classification, otherwise, directly outputting a classification result to be normal by an activation module, namely when the approximate variance Y is smaller than or equal to a threshold Th, the current heart beat signal is a normal heart beat, and directly obtaining the classification result without the neural network classifier.
When the neural network classifier detects that the currently received heartbeat signal is the second heartbeat signal (differential heartbeat signal)HB diff ) Then, the neural network classifier reads the differential neural network model M2 from the differential network memory as the current classification model to perform recognition classification, and obtains the second classification result (differential heartbeat signal)HB diff The classification result) and simultaneously sends the second classification result to the template updating module.
The template updating module obtains a differential heartbeat signalHB diff After the classification result of (2), if the result is abnormal, no processing is performed. And if the current heartbeat is normal, calculating the weighting update coefficient a of the current heartbeat by the formula (6):
wherein, threshold represents a threshold considered to be set, softmax [ normal ] is a probability value obtained by passing a calculation result of the neural network (the differential neural network model M2) through a softmax function and judged as a normal heartbeat, and similarly, softmax [ abnormal ] is a probability value obtained by passing the neural network (the differential neural network model M2) and judged as an abnormal heartbeat.
Template update module based on current heart beat templateHB temp Current differential cardiac signalHB diff And a weighting update coefficient a, calculating a new heart beat template according to formula (7)
And stored in the heart beat template memory.
In order to further improve the accuracy of classification, as a possible implementation manner, in the embodiment of the present invention, the threshold Th is further adaptively updated by the following different means:
the initial value of the threshold Th may be set to a relatively low value so that most of the cardiac signal at the beginningHB in The classification result of (2) is obtained by a neural network classifier.
In addition, when the activation module detects that the times of not waking up the neural network classifier exceed the frequency threshold, the activation module directly wakes up the neural network classifier and sends the last differential heartbeat signalHB diff To judge the corresponding heart beatSignalHB in Whether it is true normal. If classified as abnormal heart beat, the threshold Th is reduced by a certain proportion.
In addition, the activation module adjusts the threshold Th according to the following adaptive threshold adjustment method: if the neural network classification result is normal, the threshold Th is raised by formula (8), and if the result is abnormal, formula (9) is used. Where c denotes a confidence factor, and β and γ are first and second coefficients set in advance.
Because the heartbeat differential data set which can be directly used for training is difficult to obtain, in order to enable the heartbeat differential coding technology to be used on the existing complete heartbeat data set without separately collecting a special heartbeat differential data set, in the embodiment of the invention, for a given classification neural network model, the neural network training mode based on heartbeat differential coding provided by the embodiment of the invention is preferably adopted, and further, a trained non-differential neural network model M1 and a trained differential neural network model M2 are obtained.
The neural network training mode based on the heartbeat differential coding is as follows:
(1) the method continues with (2) by first dividing the continuous ECG recording into a plurality of heartbeat waveforms based on the location of a designated feature (preferably the R peak, i.e., the highest point of a single beat rhythm of the cardiac signal) in the patient ECG recording database to obtain a complete heartbeat data set D1, and into a training set and a verification set.
(2) And (3) training the network parameters of the classification neural network model of the non-differential neural network model M1 by using the training set of the data set D1, selecting the classification neural network model with the optimal accuracy by using the verification set, finally obtaining the non-differential neural network model M1, and continuing to execute the step (3).
(3) The first N beat heart signals of each patient ECG recording in data set D1 are fed into model M1 for classification until K beats are obtained and recognized as normal by model M1, and then the initial beat is calculated based on formula (1)Initial beat templateHB temp The value of K is freely definable, and execution continues with (4).
(4) Cardiac beat of ECG recording of each patient in data set D1HB in Are sequentially sent into a model M1 for classification, inHB in Obtaining the differential heartbeat signal of the heartbeat according to the formula (2) before being sent to the model M1 for classificationHB diff And after the classification result is obtained, updating the heart beat template according to the formula (7), switching to the next patient in the data set D1 and repeating the steps (3) and (4) each time the current patient is processed, finally obtaining a heart beat differential data set D2, dividing the heart beat differential data set D2 into a training set and a verification set, and continuing to execute the step (5).
(5) And training the network parameters of the classified neural network model of the differential neural network model M2 by using the training set of the heartbeat differential data set D2, selecting the classified neural network model with the optimal accuracy by using the verification set, and finally obtaining the differential neural network model M2.
As a possible implementation manner, the work flows of the initialization stage and the working stage of the artificial intelligence processing circuit for cardiac signals based on cardiac beat differential coding provided in the embodiment of the present invention are respectively shown in fig. 2 and 3:
and S0, no matter what stage the heart beat is, the preprocessing module detects the heart beats from the continuous electrocardiosignals, extracts the heart beat signals one by one according to the position of the heart beat and the certain window length, and stores the heart beat signals in the heart beat memory.
S1, in the initialization stage, the differential encoder takes out the complete heartbeat from the heartbeat memory and directly sends the complete heartbeat to the neural network classifier. The neural network classifier takes out the non-differential network model M1 from the non-differential network memory for identification and classification. The template updating module carries out the identified K normal heartbeatsHB normal Calculating the average value as the initial heart beat templateHB temp And will beHB temp And storing the heart beat template into a heart beat template memory.
S21, in the working stage, each time a heartbeat signal is obtained, the differential encoder finishes extracting the heartbeat signal from the heartbeat memoryWhole heart beat signalHB in According toHB in And current heart beat templateHB temp Calculating a differential cardiac signal from the difference ofHB diff And will beHB diff To the activation module.
S22, second step of working stage: the activation module will differentially beat the signalHB diff Divided into M sections, each section containing N sampling points s1~sN. For the mth segment, taking the first N/2 points to calculate the mean value according to the formula (3)X mAccording to formula (4) from the mean valueX mAnd post N/2 point calculationY m And finally, averaging the M sections of signals through a formula (5) to obtain an approximate variance Y for measuring the similarity of the current heart beat and the heart beat template.
The activation module compares Y with a threshold Th, activates the neural network if Y is greater than the threshold, and differentiates the cardiac beat signalHB diff Sending the data to a neural network classifier for further classification, and continuing to execute the step S23; otherwise, the heart beat is classified as normal heart beat, and the activation module outputs the classification result as normal. If the heart beat is the first heart beat, the correspondence is compared with the initial threshold value. The initial threshold setting is low, so most heartbeats will be classified by the neural network classifier. In addition, when the activation module detects that the accumulated times of not waking up the neural network classifier exceeds the frequency threshold, the neural network classifier is directly woken up and a last differential heart beat signal is sent to judge whether the heart beat is normal or not. If classified as abnormal heart beat, the threshold Th is reduced by a certain proportion.
S23, a third step of the working stage: the neural network classifier receives the differential heartbeat signalHB diff The differential network model M2 is then fetched from the differential network memory, and is paired with the differential network memoryHB diff And classifying and outputting a classification result.
S24, fourth step of working stage: the template update module is obtainingHB diff After the classification result of (2), if the result is abnormal, no processing is performed. And if normal, the public is passedEquation (6) calculates the weighting update coefficient a of the current heartbeat. In addition, the activation module adjusts the threshold Th in step S22 according to the following adaptive threshold adjustment method: if the neural network classification result is normal, the threshold Th is raised by formula (8), and if the result is abnormal, formula (9) is used.
S25, and finally, in the fifth working stage: and (5) updating the template. The template update module uses the current heart beat templateHB temp Current differential cardiac signalHB diff And weighting update coefficient a obtained by softmax weighting, and calculating a new heart beat template according to formula (7)
And stored in the heart beat template memory.
S31, in the working phase, the steps S21, S22, S23, S24 and S25 are circulated once every time the heartbeat signal is input.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
What has been described above are merely some embodiments of the present invention. It will be apparent to those skilled in the art that various changes and modifications can be made without departing from the inventive concept thereof, and these changes and modifications can be made without departing from the spirit and scope of the invention.
Claims (6)
1. An artificial intelligence processing circuit of electrocardiosignals based on beat differential coding is characterized by comprising: the device comprises a preprocessing module, a heartbeat memory, a heartbeat template memory, a differential coding module, an activation module, a neural network classifier, a differential network memory, a non-differential network memory and a template updating module, wherein the preprocessing module is connected with the differential coding module through the heartbeat memory, the differential coding module is sequentially connected with the activation module and the neural network classifier, the differential network memory and the non-differential network memory are connected with the neural network classifier, the neural network classifier is connected with the heartbeat template memory through the template updating module, and the heartbeat template memory is connected with the differential coding module;
the preprocessing module is used for preprocessing the input electrocardiosignals, detecting the heartbeat signals, extracting single heartbeat signals and storing the extracted heartbeat signals to the heartbeat memory;
the non-differential network memory is used for storing a pre-trained non-differential neural network model M1, and the differential network memory is used for storing a pre-trained differential neural network model M2; the non-differential neural network model M1 and the differential neural network model M2 are used for carrying out two classification recognition processing of normal heartbeat and abnormal heartbeat on the input heartbeat signal;
the differential encoder comprises two working modes, which are respectively:
in the first working mode, the differential encoder reads a heartbeat signal from the heartbeat memory as a first heartbeat signal and sends the first heartbeat signal to the neural network classifier, and when the neural network classifier detects that the currently received heartbeat signal is the first heartbeat signal, the neural network classifier reads a non-differential neural network model M1 from the non-differential network memory as a current classification model for identification and classification to obtain a first classification result and sends the first classification result to the template updating module; the template updating module takes the average value of normal heartbeats in the first classification results as an initial heart beat template and stores the initial heart beat template into a heart beat template memory based on the obtained plurality of first classification results;
in the second working mode, the differential encoder reads the heart beat signal currently obtained by the preprocessing module from the heart beat memory in real time, obtains a differential heart beat signal based on the difference value between the current heart beat signal and the current heart beat template, and sends the differential heart beat signal to the activation module; the activation module is used for calculating the variance of the differential heartbeat signal, if the variance of the current differential heartbeat signal is larger than a threshold Th, the neural network classifier is awakened, and the current differential heartbeat signal is sent to the neural network classifier; if the variance of the current differential heartbeat signal is less than or equal to the threshold Th, the classification result of the current heartbeat signal is normal heartbeat, and the classification result of the current heartbeat signal is directly output;
when the neural network classifier receives the differential heartbeat signal, reading a differential neural network model M2 from a differential network memory as a current classification model for identification and classification, obtaining a second classification result and sending the second classification result to a template updating module;
when the template updating module detects that the current second classification result is normal heartbeat, calculating a weighting updating coefficient a of the current differential heartbeat signal according to a configured calculation strategy, obtaining an updating amount according to the product of the weighting updating coefficient a and the current differential heartbeat signal, and updating the heartbeat template stored in the heartbeat template memory, wherein the updated heartbeat template is as follows: the sum of the pre-update heart beat template and the update amount.
2. The artificial intelligence processing circuit for electrocardiosignals based on the beat differential coding as claimed in claim 1, wherein the weighting update coefficient a is: the current differential heartbeat signal output by the differential neural network model M2 is judged to be one half of the difference between the probability value of normal heartbeat and the probability value of abnormal heartbeat, and then multiplied by a preset threshold value threshold.
3. The artificial intelligence processing circuit for cardiac signals based on cardiac beat differential coding according to claim 1, wherein the specific way of calculating the variance of the differential cardiac beat signal by the activation module is:
uniformly dividing the differential heart beat signal into M sections, wherein M is an integer greater than 1; and defining the number of sampling points contained in each section as N;
for the mth section differential heart beat signal, the first N/2 sampling points are taken to calculate the mean valueX mWherein M =1,2, …, M;
respectively calculating each sampling point and the mean value of the last N/2 sampling points of the mth section differential cardiac signalX mAnd the cumulative sum of the absolute differences is recorded asY m ；
Based on M-segment difference cardiodyne signalsY m The mean of (a) yields an approximate variance and serves as the variance of the differential heartbeat signal.
4. The artificial intelligence heart signal processing circuit based on beat differential encoding as claimed in claim 1, wherein said activation module further comprises: when the accumulated times of not waking up the neural network classifier exceeds a frequency threshold value, directly waking up the neural network classifier, sending a latest differential heartbeat signal to the neural network classifier, waiting for the neural network classifier to feed back a classification result, and if the differential heartbeat signal is abnormal heartbeat, reducing the threshold value Th according to a specified proportion.
5. The artificial intelligence processing circuit for electrocardiosignals based on the beat differential coding as claimed in any one of claims 1 to 4, wherein the activation module performs adaptive threshold adjustment on the threshold Th according to the classification result currently output by the neural network classifier in real time: if the classification result is normal heartbeat, updating the threshold Th according to a formula Th = Th + betax (Y-Th) x c; if the classification result is abnormal heartbeat, updating the threshold Th according to a formula Th = gamma multiplied by Y multiplied by c;
where Y denotes a variance of the differential heartbeat signal, c denotes a preset confidence factor, and β and γ denote preset first and second coefficients, respectively.
6. The artificial intelligence processing circuit for electrocardiosignals based on the beat differential coding as claimed in claim 1, wherein the training modes of the non-differential neural network model M1 and the differential neural network model M2 comprise:
(1) dividing continuous electrocardiosignals into a plurality of heart beat waveforms based on the designated characteristic positions of the electrocardiosignals in the patient electrocardiosignal recording database, thereby obtaining a heart beat data set D1, and dividing the heart beat data set D into a training set and a verification set according to a designated proportion;
(2) training network parameters of a classification neural network model of the non-differential neural network model M1 by using a training set of a heartbeat data set D1, and selecting the classification neural network model with the optimal accuracy rate through a verification set of a heartbeat data set D1 to obtain a non-differential neural network model M1;
(3) sending the first N 'heart beating signals recorded by the electrocardiosignals of each patient in the heart beating data set D1 into a non-differential neural network model M1 for classification, counting the number of normal heart beats as a classification result, and obtaining an initial heart beating template based on the mean value of K normal heart beats when K heart beats are obtained, wherein N' and K are positive integers more than 1;
(4) sequentially traversing the heartbeat signal of the electrocardiosignals of each patient in the heartbeat data set D1HB in Based on the current heartbeat signalHB in Obtaining a differential heartbeat signal by the difference between the current heartbeat template and the current heartbeat templateHB diff Sequentially sending the heart beat data to a non-differential neural network model M1 for classification, and updating the heart beat template based on the currently obtained classification result;
based on all differential heartbeat signalsHB diff Obtaining a heart beat differential data set D2, and dividing the heart beat differential data set D2 into a training set and a verification set according to a specified proportion;
(5) and training the network parameters of the classified neural network model of the differential neural network model M2 by using the training set of the heartbeat differential data set D2, and selecting the classified neural network model with the optimal accuracy rate by using the verification set of the heartbeat differential data set D2 to obtain the differential neural network model M2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210308702.0A CN114469127B (en) | 2022-03-28 | 2022-03-28 | Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210308702.0A CN114469127B (en) | 2022-03-28 | 2022-03-28 | Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114469127A true CN114469127A (en) | 2022-05-13 |
| CN114469127B CN114469127B (en) | 2022-07-29 |
Family
ID=81488572
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210308702.0A Expired - Fee Related CN114469127B (en) | 2022-03-28 | 2022-03-28 | Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114469127B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103549950A (en) * | 2013-11-19 | 2014-02-05 | 上海理工大学 | Improved difference threshold detection algorithm for mobile ECG (electrocardiogram) monitoring |
| CN104207769A (en) * | 2014-08-27 | 2014-12-17 | 电子科技大学 | Electrocardiosignal detection system |
| US20160120430A1 (en) * | 2014-10-29 | 2016-05-05 | Khalifa University of Science, Technology, and Research | Medical device for detecting a ventricular arrhythmia event |
| CN108714026A (en) * | 2018-03-27 | 2018-10-30 | 杭州电子科技大学 | The fine granularity electrocardiosignal sorting technique merged based on depth convolutional neural networks and on-line decision |
| CN108888259A (en) * | 2018-05-21 | 2018-11-27 | 南京大学 | A kind of real-time QRS wave detection method of electrocardiosignal |
| CN112971800A (en) * | 2021-01-27 | 2021-06-18 | 东南大学 | Deep learning-based electrocardio premature beat signal processing method, module and device |
| EP3878361A1 (en) * | 2020-03-12 | 2021-09-15 | Siemens Healthcare GmbH | Method and device for determining a cardiac phase in magnet resonance imaging |
| US20210369174A1 (en) * | 2020-05-27 | 2021-12-02 | Biosense Webster (Israel) Ltd. | Automatic detection of cardiac structures in cardiac mapping |
-
2022
- 2022-03-28 CN CN202210308702.0A patent/CN114469127B/en not_active Expired - Fee Related
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103549950A (en) * | 2013-11-19 | 2014-02-05 | 上海理工大学 | Improved difference threshold detection algorithm for mobile ECG (electrocardiogram) monitoring |
| CN104207769A (en) * | 2014-08-27 | 2014-12-17 | 电子科技大学 | Electrocardiosignal detection system |
| US20160120430A1 (en) * | 2014-10-29 | 2016-05-05 | Khalifa University of Science, Technology, and Research | Medical device for detecting a ventricular arrhythmia event |
| CN108714026A (en) * | 2018-03-27 | 2018-10-30 | 杭州电子科技大学 | The fine granularity electrocardiosignal sorting technique merged based on depth convolutional neural networks and on-line decision |
| CN108888259A (en) * | 2018-05-21 | 2018-11-27 | 南京大学 | A kind of real-time QRS wave detection method of electrocardiosignal |
| EP3878361A1 (en) * | 2020-03-12 | 2021-09-15 | Siemens Healthcare GmbH | Method and device for determining a cardiac phase in magnet resonance imaging |
| US20210369174A1 (en) * | 2020-05-27 | 2021-12-02 | Biosense Webster (Israel) Ltd. | Automatic detection of cardiac structures in cardiac mapping |
| CN112971800A (en) * | 2021-01-27 | 2021-06-18 | 东南大学 | Deep learning-based electrocardio premature beat signal processing method, module and device |
Non-Patent Citations (1)
| Title |
|---|
| 张俊飞等: "基于深度学习的ECG心拍数据分类设计", 《自动化与仪器仪表》 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114469127B (en) | 2022-07-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106725428B (en) | Electrocardiosignal classification method and device | |
| US8290575B2 (en) | Biomedical signal morphology analysis method | |
| CN110623658B (en) | Signal processing method, signal processing apparatus, medical device, and storage medium | |
| Xia et al. | A novel wearable electrocardiogram classification system using convolutional neural networks and active learning | |
| EP1750580A2 (en) | Signal analysis method | |
| CN110897633A (en) | Electrocardiosignal segmentation method based on deep neural network | |
| CN111067503A (en) | Sleep staging method based on heart rate variability | |
| CN112617850B (en) | Premature beat and heart beat detection system for electrocardiosignals | |
| CN114052744A (en) | Electrocardiosignal classification method based on pulse neural network | |
| CN109887595A (en) | Heartbeat anomalous identification algorithm based on depth learning technology | |
| CN106730232A (en) | A kind of intelligent awakening method and system | |
| CN114469127B (en) | Electrocardiosignal artificial intelligence processing circuit based on heart beat differential coding | |
| Sakhare et al. | An approach for ECG beats classification using adaptive neuro fuzzy inference system | |
| CN115470832B (en) | Electrocardiosignal data processing method based on block chain | |
| CN116172572A (en) | SHAP value weighted sum segmentation HDBSCAN-based heart beat clustering method and system | |
| Surrel et al. | Design of ultra-low-power smart wearable systems | |
| CN110786847B (en) | Electrocardiogram signal library building method and analysis method | |
| US20200345911A1 (en) | Predictive qrs detection and r-to-r timing systems and methods | |
| CN112022149A (en) | Atrial fibrillation detection method based on electrocardiosignals | |
| US20220288380A1 (en) | Predictive qrs detection and r-to-r timing systems and methods | |
| CN111685754B (en) | Heart rate calculation method and system for wearable ECG acquisition equipment | |
| Surapaneni et al. | Graph Signal Processing Based Classification of Noisy and Clean PPG Signals Using Machine Learning Classifiers for Intelligent Health Monitor | |
| Li et al. | ECG segmentation in a body sensor network using Hidden Markov Models | |
| CN114869226A (en) | Sleep staging method based on cardiac shock signal | |
| CN117493768A (en) | Early ventricular recognition method based on improved residual neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220729 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |