CN113951893A

CN113951893A - Multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge

Info

Publication number: CN113951893A
Application number: CN202111460169.1A
Authority: CN
Inventors: 师丽; 韩闯; 王松伟; 王治忠
Original assignee: Tsinghua University; Zhengzhou University
Current assignee: Tsinghua University; Zhengzhou University
Priority date: 2021-12-02
Filing date: 2021-12-02
Publication date: 2022-01-21
Anticipated expiration: 2041-12-02
Also published as: CN113951893B

Abstract

The invention relates to the technical field of an extraction method of electrocardiosignal extraction points, and discloses a multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge, wherein firstly, a multi-lead electrocardiosignal acquisition module is used for extracting 12-lead electrocardiosignals; secondly, the feature point extraction module extracts morphological features of the heart beat and strong time sequence correlation features of sampling moments through a Convolutional Neural Network (CNN) and a long-short term memory network (LSTM) based on a U-net frame, strengthens finer features of each moment of a waveform through fusion of bottom layer information and high layer information, and then extracts feature points through a fixed threshold method; finally, the feature point position correction module further improves the feature point extraction precision through a multi-lead mutual reference method based on electrophysiological knowledge and a dynamic threshold self-adaptive adjustment strategy, and can remarkably reduce the missed diagnosis rate and the misdiagnosis rate of feature point extraction.

Description

Multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge

Technical Field

The invention relates to the technical field of an electrocardiosignal characteristic point extraction method, in particular to a multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge.

Background

The electrocardiosignal is a weak electrophysiological signal generated by the heart, can evaluate the heart state and diagnose related cardiovascular diseases, and mainly comprises four waves of P wave, QRS wave, T wave, U wave, four segments of PR interval, QRS time, ST segment and QT interval. The key characteristic point extraction and positioning technology is the basis of electrocardio diagnosis and analysis, and has important significance for diagnosing arrhythmia, myocardial infarction, atrioventricular hypertrophy and other cardiovascular diseases. The existing method for peak points of an electrocardio QRS wave group comprises filtering, difference threshold, mathematical morphology, wavelet transformation, Hilbert transformation, empirical mode decomposition, Hilbert-yellow transformation, machine learning based on artificial feature extraction, deep learning and the like; the identification of the QRS wave starting and stopping points is generally to set a time window on the basis of the identification of the peak point and search the starting and stopping points of the QRS wave based on a slope threshold method, and researchers also identify the QRS wave starting and stopping points by a local transformation method, a least square fitting method, machine learning and other methods, and a wavelet analysis method, a trapezoidal area method, a time window searching method and the like are used for extracting the T wave peak point and the stopping points.

Most of the methods need to rely on experience parameters and various characteristics of manual design, and lack of generalization and robustness, while the existing detection method based on deep learning can automatically extract the characteristics of waveforms at various moments, the identification accuracy rate is still to be improved for the situations that electrocardiosignals contain specific QRS complexes, ST-T sections with complex forms, malignant arrhythmia and the like, and the missed diagnosis rate and the misdiagnosis rate are high. Therefore, the existing electrocardiosignal extraction has the defects of low accuracy, high missed diagnosis rate, high misdiagnosis rate and the like.

Disclosure of Invention

Technical problem to be solved

Aiming at the defects of the prior art, the invention provides the method for extracting the multi-lead electrocardiosignal characteristic points by combining deep learning and electrophysiological knowledge, which has the advantages of higher accuracy, low missed diagnosis rate and misdiagnosis rate and the like and solves the problems in the background art.

(II) technical scheme

In order to achieve the purposes of higher precision and low missed diagnosis rate and misdiagnosis rate, the invention provides the following technical scheme:

the invention provides a multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge, which comprises the following steps:

(1) the multi-lead electrocardiosignal acquisition module: acquiring multi-lead electrocardiosignals through a signal acquisition sensor, an isolation circuit, an amplification circuit, a filter circuit and an analog/digital conversion circuit, and separating the multi-lead electrocardiosignals into twelve-lead signals;

(2) a feature point extraction module: inputting the preprocessed electrocardio data into a pre-trained model to detect electrocardio signal characteristic points, acquiring the probability that each time sampling point belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave ending point, a T wave peak point and a T wave ending point, and extracting the characteristic points based on a fixed threshold method;

(3) a characteristic point position correction module: and (3) further correcting the positions of a QRS wave starting and stopping point, a QRS wave peak point, a T wave peak point and a T wave stopping point on the basis of a dynamic threshold self-adaptive adjustment strategy and a multi-lead mutual reference method based on electrophysiological knowledge on the basis of the step (2).

Preferably, the feature point extraction module includes a constructed data set, the marked multi-lead electrocardiographic data is subjected to baseline interference removal and high-frequency and low-frequency noise removal through a db6 wavelet-based discrete wavelet transform, a 0.5-40 Hz third-order butterworth band-pass filter, the length of each lead electrocardiographic data is intercepted to be N, 80% of the data is a training set, 20% of the data is a testing set, and after the construction of the data set is completed, if a certain sampling point is marked as a QRS wave starting point, the sampling points of the first 0.075s and the second 0.075s of the sampling point are also considered as QRS wave starting points, that is, the sampling points have the same label.

Preferably, the feature point extraction module performs a link of model establishment and training after the data set is constructed and processed, and the link of model establishment and training adopts the following two steps: establishing a convolution neural network based on a U-net framework; and establishing a long-short term memory network based on a U-net framework and training the model.

Preferably, the convolutional neural network based on U-net framework (U-net-ECGCNN) has 19 layers, which comprise an encoding module, an underlying module, a decoding module and a Softmax layer;

the coding module comprises 6 layers in total and consists of 6 one-dimensional convolutional layers (1D-C), 3 batch normalization layers (BN), 3 Relu layers and three one-dimensional down-sampling layers (1D-P), wherein the length of a convolutional kernel is S multiplied by 1, the numerical value of S in the first two layers is 31, then each two layers are reduced by 6, the number of the convolutional kernels is 16 multiplied by 2K, the value of K is equal to 0 at first, and the numerical value of each two convolutional layers is increased by 1. The down-sampling factor of each down-sampling layer is 2;

the bottom layer module comprises 2 layers in total and consists of two 1D-C layers, 1 BN layer and a Relu layer; wherein the length and number of convolution kernels are 13 and 128 respectively;

the decoding module consists of 10 one-dimensional convolutional layers, 6 BN layers, 6 Relu layers, three one-dimensional upsampling (1D-U) layers and three skip layer link (SC) layers. The number of convolutional kernels in the convolutional layer is 16 × 2K, the value of K is initially equal to 2, and every two convolutional layer values are reduced by 1. The upsampling factor of each upsampling layer is 2; wherein the 1D-U can recover the detail information of each sampling point, and the parameters thereof correspond to the coding module. The SC may provide additional information lost during the 1D-U phase;

the Softmax layer outputs the probability that each sampling point belongs to the starting point, the peak point, the ending point of the QRS wave and the peak point and the ending point of the T wave respectively.

Preferably, the U-net framework-based long-short term memory network comprises an encoding module, a bottom layer module, a decoding module and a Softmax layer, and is different from the U-net framework-based convolutional neural network in that the bottom layer module is replaced by two LSTM layers for extracting strong time sequence correlation characteristics at each sampling time, and the number of hidden layer nodes in the two LSTM layers is 128.

Preferably, the model loss function of the model training is:

where n is the number of batch samples per input network and J is the number of categories. z represents the probability that the sampling time T belongs to the class K, K belongs to K, T belongs to 0,1,2, …, N, T belongs to T, and y is a sample label value, the model is trained by a random gradient descent (SGD) optimizer, and the model loss function is minimized through gradual iteration. With the momentum, learning rate and batch size superparameters being 0.9, 0.005 and 64, respectively. And when the loss function value is not reduced after the U-net-ECGCNN model and the U-net-ECGLSTM model are continuously trained for 20 rounds respectively, stopping training and respectively storing the model M and the model N at the moment.

Preferably, the preprocessed input signal { x (T), T ═ 1,2, … N } is input into the U-net-ECGCNN model and the U-net-ECGLSTM model, and probability matrices P and Q of sampling points at each time belonging to the QRS wave start point, QRS wave peak point, QRS wave end point, T wave peak point, and T wave end point are obtained respectively, where the expressions are:

wherein P is_N1，P_N2，P_N3，P_N4，P_N5Representing the probability that the Nth sampling point based on the U-net-ECGCNN model respectively belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point; wherein Q_N1，Q_N2，Q_N3，Q_N4，Q_N5And expressing the probability that the Nth sampling point belongs to the QRS wave starting point, the QRS wave peak point, the QRS wave termination point, the T wave peak point and the T wave termination point respectively based on the U-net-ECGLSTM model.

Preferably, the probability matrix V that the sampling point at each time is the corresponding feature point is obtained based on ensemble learning, and the expression is as follows:

wherein:

and converting it into a matrix consisting of 0 to 1, whose expression is:

if the probability value V_j1Greater than thr1, then X_j1Equal to 1, otherwise X_j1Equal to 0. For the detection of the starting point of the QRS wave, the probability matrix of each sampling point belonging to the starting point of the QRS wave is W1 ═ W₁,W₂,…W_j...W_N]The matrix consisting of 0 and 1 after conversion is X1 ═ X₁,X₂,…X_N]。

The specific steps for finding the position of the starting point of the QRS wave are as follows:

finding a first position which is equal to 1 in X1, recording the position as X1, searching from the position of X1 until the position of the next X2 is found to be equal to 1, recording the position of X2+1 as 0, and recording the position X2 at the moment, wherein the middle position from X1 to X2 is the starting point p1 of the first QRS wave;

continuing searching backwards from the position x2+1, and recording the positions of all QRS wave starting points;

the QRS start-stop point position p is [ p ]₁,p₂,…,p_j,…p_y]Wherein y represents the number of QRS wave starting points based on a fixed threshold method; j denotes an index value.

Preferably, the feature point correcting module corrects the position of the feature point by a multi-lead cross-reference method based on electrophysiological knowledge and corrects the position of the feature point by a dynamic threshold adaptive adjustment strategy, wherein a single-lead signal x (t) of the M-lead electrocardiograph signal in step (1) is obtained, and a probability matrix of each sampling point belonging to a QRS wave starting point in the method based on step (2) is W1 ═ W₁,W₂,…W_j...W_N]And (3) applying the method in the step (2) to all the rest leads of the M-lead electrocardiosignals, wherein the total probability matrix of each sampling point of the M-lead electrocardiosignals belonging to the QRS wave initial point is as follows:

W_INthe probability that the Nth sampling point on the I-th lead belongs to the QRS initial point is represented, I is a lead index value, the probability mean value that the sampling point belongs to the QRS wave initial point on each lead is calculated by a multi-lead mutual reference method based on electrophysiological knowledge, the probability mean value is used as the probability that the sampling point belongs to the QRS initial point on the lead, and after the probability matrix that each sampling point belongs to the QRS wave initial point is applied to the N sampling points:

W＝[W₁,W₂,W₃,....,W_I,..W_N]

wherein

Wherein W_NIndicating the probability that the nth sample belongs to the QRS starting point.

8. Preferably, the position of the feature point is corrected based on a dynamic threshold adaptive adjustment strategy, and the position p of the starting point of the QRS wave is [ p ]₁,p₂,…,p_j,…p_y]Detecting all elements of the matrix p one by setting different thresholds, the detection of which isIf not, the QRS wave starting point is judged, and the result is stored in the matrix QRS, and the specific steps are as follows:

step a, if p [1]]≧ Fs, indicating that one or more QRS wave start points are missed, the new probability matrix Y ═ W is used₁,W₂,…W_p[0]+1/5*Fs]And detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr2 is equal to thr1-0.1, and storing the detection result in the matrix QRS. If QRS is empty, lowering threshold value thr 3-thr 2-0.1 until QRS is not empty, if QRS is not empty, saving QRS and transferring to step b; if p [1]]＜Fs，p[1]Marking as the starting point of the first QRS wave, and p1]And (4) storing the QRS matrix, and then, turning to the step b to carry out the next step.

Step b, detecting p 2 in matrix p one by one]To p [ y-1]All elements in the middle, detecting whether the element is the starting point of the QRS wave, and recording p [ j]Is the j-1 st point to be detected, if (pj)]-QRS[-1]) Not less than 1.2 xFs, then pj]And QRS-1]And missing the QRS wave starting point. Then the new probability matrix Z ═ W is used_{QRS[-1]+1/5*Fs},…W_p[j]-1/5*Fs]And detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr3 is equal to thr1-0.1, and storing the detection result in the matrix QRS. If QRS is empty, lowering threshold to thr 4-thr 3-0.1 until QRS is not empty, if QRS is not empty, saving QRS and going to step c; if (p [ j)]-QRS[-1])＜1.2×Fs,p[j]For the starting point of QRS wave, p [ j ] is]And storing the QRS matrix into a QRS matrix, and then turning to the step c to carry out the next step.

Step c, if p [ -1]]Less than or equal to N-Fs, then p-1]And the QRS wave starting point is missed to be detected between N. Then the new probability matrix T ═ W is used_P[-1]+1/5*Fs,…W_N]And detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr5 is equal to thr1-0.1, and storing the detection result in the QRS matrix. If QRS is empty, lowering threshold value thr 6-thr 5-0.1 until QRS is not empty, if QRS is not empty, storing QRS; if p is [ -1]]＞N-Fs,p[j]For the starting point of QRS wave, p [ j ] is]Save to the matrix QRS. Last QRS ═ QRS₁,QRS₂,…,QRS_j,…,QRS_z]Wherein z represents the number of QRS wave starting points based on the dynamic threshold self-adaptive adjustment method; j denotes an index value. And extracting a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point according to the same method.

(III) advantageous effects

Compared with the prior art, the invention provides a multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge, which has the following beneficial effects:

1. according to the method for extracting the multi-lead electrocardiosignal characteristic points by combining deep learning and electrophysiological knowledge, morphological characteristics of electrocardiosignal heart beat and strong time sequence correlation characteristics of sampling time can be represented respectively through CNN and LSTM networks based on a U-net frame, and finer characteristics of each time of a waveform are enhanced through fusion of bottom layer information and high layer information, so that the purpose of providing precision is achieved, through setting of self-adaptive adjustment strategy correction based on dynamic thresholds, thresholds are gradually reduced, related characteristic points are searched again, the missed detection rate can be remarkably reduced, and the purpose of reducing the missed diagnosis rate is achieved.

2. According to the method for extracting the multi-lead electrocardiosignal characteristic points by combining deep learning and electrophysiological knowledge, the mean value of the probability that each sampling point of a plurality of leads belongs to the corresponding characteristic point is calculated through the setting of a multi-lead mutual reference method, so that the accuracy of positioning the characteristic points is further improved, the characteristic points can be accurately positioned when QRS wave groups with complex shapes and ST-T sections, noise interference is large, and malignant arrhythmia events occur, and the purposes of reducing misdiagnosis rate and missed diagnosis rate are achieved.

Drawings

FIG. 1 is a block diagram of a method for extracting feature points of a multi-lead ECG signal feature point combining deep learning and electrophysiological knowledge according to the present invention;

FIG. 2 is a schematic diagram of a process of correcting positions of feature points based on a dynamic threshold adaptive adjustment strategy for a multi-lead ECG signal feature point extraction method combining deep learning and electrophysiological knowledge according to the present invention;

FIG. 3 is a schematic diagram of the extraction result of feature points on an LUDB data set by a multi-lead electrocardiosignal feature point extraction method combining deep learning and electrophysiological knowledge according to the present invention;

FIG. 4 is a schematic diagram of the extraction result of feature points on a CCDD data set of a multi-lead electrocardiosignal feature point extraction method combining deep learning and electrophysiological knowledge according to the present invention;

FIG. 5 is a schematic diagram showing the comparison of the characteristic point extraction results of different methods on a LUDB data set by a multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge provided by the present invention;

FIG. 6 is a schematic diagram showing the comparison of the feature point extraction results of different methods on a CCDD data set by a multi-lead electrocardiosignal feature point extraction method combining deep learning and electrophysiological knowledge provided by the present invention;

FIG. 7 is a schematic diagram of a model structure of a multi-lead ECG signal feature point extraction method combining deep learning and electrophysiological knowledge according to the present invention;

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to the attached drawings 1-7 in the specification, the invention provides a method for extracting multi-lead electrocardiosignal characteristic points by combining deep learning and electrophysiological knowledge, which specifically comprises the following steps:

the method comprises the following steps: the multi-lead electrocardiosignal acquisition module: acquiring multi-lead electrocardiosignals through a signal acquisition sensor, an isolation circuit, an amplification circuit, a filter circuit and an analog/digital conversion circuit, and separating the multi-lead electrocardiosignals into twelve-lead signals;

step two: a feature point extraction module: inputting preprocessed electrocardio data into a pre-trained model to detect electrocardio signal characteristic points, acquiring the probability that each moment sampling point belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave ending point, a T wave peak point and a T wave ending point, and extracting the characteristic points based on a fixed threshold method, wherein the method specifically comprises the following contents:

(1) building a data set

Preprocessing data: removing base line interference from the multi-lead electrocardiogram data with labels through discrete wavelet transformation based on db6 wavelet and removing high-frequency and low-frequency noise through a three-order Butterworth band-pass filter of 0.5-40 Hz, intercepting the length of each lead electrocardiogram data as N, using 80% of data as a training set, and using 20% of data as a test set. If the point to be detected is the initial point of the QRS wave, the label is (1,0,0,0, 0); if the QRS wave peak point is present, the label is (0,1,0,0, 0); if the QRS wave termination point is present, the label is (0,0,1,0, 0); the label is (0,0,1,0, 0); if the T wave peak value point is the T wave peak value point, the label is (0,0,0,1, 0); if the T wave termination point is the T wave termination point, the label is (0,0,0,0, 1). The dimensionality of input data is Nx 1, and the dimensionality of a data label is Nx 5;

secondly, label expansion: if a certain sampling point is marked as the starting point of the QRS wave, the sampling points of the first 0.075s and the last 0.075s of the sampling point are considered as the starting point and the ending point of the QRS wave, i.e. they have the same label

(2) Model building and training

Establishing Convolution Neural Network (CNN) based on U-net frame

The convolutional neural network based on the U-net framework (U-net-ECGCNN) has 19 layers, and comprises an encoding module, an underlying module, a decoding module and a Softmax layer;

the coding module comprises 6 layers which are composed of 6 one-dimensional convolution layers (1D-C), 3 batch normalization layers (BN), 3 Relu layers and three one-dimensional down-sampling layers (1D-P), wherein the length of a convolution kernel is S multiplied by 1, the numerical value of S in the first two layers is 31, then each two layers are reduced by 6, the number of the convolution kernels is 16 multiplied by 2K, the value of K is equal to 0 at first, the numerical values of every two convolution layers are increased by 1, and the down-sampling factor of each down-sampling layer is 2; the bottom layer module comprises 2 layers in total and consists of two 1D-C layers, 1 BN layer and a Relu layer; wherein the length and number of convolution kernels are 13 and 128 respectively;

Establishing a long-short term memory network (LSTM) based on a U-net framework,

the convolutional neural network (U-net-ECGLSTM) based on the U-net framework comprises a coding module, a bottom layer module, a decoding module and a Softmax layer, and is different from the U-net-ECGCNN in that the bottom layer module is replaced by two LSTM layers and used for extracting strong time sequence correlation characteristics of each sampling moment. The number of hidden nodes in the two LSTM layers is 128.

Model training

The model loss function is:

The QRS wave and the ST-T section are detected in the step two, and the probability that the sampling point at each moment is the corresponding characteristic point is obtained by the following steps:

(1) inputting an input signal { x (T), T1, 2, … N } into a U-net-ECGCNN model and a U-net-ECGLSTM model, and respectively obtaining probability matrixes P and Q of sampling points at each moment belonging to a QRS wave starting point, a QRS wave peak point, a QRS wave ending point, a T wave peak point and a T wave ending point, wherein the expressions are respectively as follows:

the PN1, the PN2, the PN3, the PN4 and the PN5 represent the probabilities that the Nth sampling point respectively belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave ending point, a T wave peak point and a T wave ending point based on a U-net-ECGCNN model; wherein QN1, QN2, QN3, QN4 and QN5 represent the probability that the Nth sampling point respectively belongs to the QRS wave start and stop point, the QRS wave peak point, the QRS wave stop point, the T wave peak point and the T wave stop point based on the U-net-ECGLSTM model.

(2) Acquiring a probability matrix V of which the sampling point at each moment is a corresponding characteristic point based on ensemble learning, wherein the expression is as follows:

wherein:

and (3) converting the characteristic matrix V obtained in the step two into a matrix X consisting of 0 and 1, wherein the expression is as follows:

specifically, the method comprises the following steps: if the probability value V_j1Greater than thr1, then X_j1Equal to 1, otherwise X_j1Equal to 0. For the detection of the starting point of the QRS wave, each sampling point belongs toThe probability matrix of the starting point of the QRS wave is W1 ═ W₁,W₂,…W_j...W_N]The matrix consisting of 0 and 1 after conversion is X1 ═ X₁,X₂,…X_N]。

Step three: and step three, further accurately positioning a QRS wave starting and stopping point, a QRS wave peak point, a T wave peak point and a T wave stopping point by the multi-lead mutual reference method based on the dynamic threshold self-adaptive adjustment strategy and the electrophysiological knowledge.

The method comprises the following steps:

1, correcting the positions of the characteristic points by a multi-lead mutual reference method based on electrophysiological knowledge, comprising the following steps:

according to electrophysiological knowledge: the QRS wave starting point, the QRS wave termination point and the T wave termination point corresponding to each lead are basically the same in position, and the accuracy rate of feature point detection can be further improved by combining the information of each lead. Wherein, in the first step, a certain single lead signal X (t) of the M lead electrocardiosignals, and the probability matrix of each sampling point belonging to the QRS wave starting point is W1 ═ based on the method in the second step₁,W₂,…W_j...W_N]And (3) applying the method in the step (2) to all the rest leads of the M-lead electrocardiosignals, wherein the total probability matrix of each sampling point of the M-lead electrocardiosignals belonging to the QRS wave initial point is as follows:

W＝[W₁,W₂,W₃,....,W_I,..W_N]

wherein

2, the step of adaptively adjusting the position of the strategy correction feature point based on the dynamic threshold is as follows:

aiming at the QRS wave starting point position p ═ p obtained by a fixed threshold value method in the step two₁,p₂,…,p_j,…p_y]

(1) And step a, if p [1] ≧ Fs indicates that one or more QRS wave starting points are missed, detecting the QRS wave starting points in the range by using a new probability matrix Y [ W1, W2, … Wp [0] +1/5 × Fs ], detecting according to the fixed threshold method in step two, and storing the detection result into the QRS matrix at the moment that the threshold thr2 ═ thr 1-0.1. If QRS is empty, lowering threshold value thr 3-thr 2-0.1 until QRS is not empty, if QRS is not empty, saving QRS and transferring to step b; if p1 < Fs, p1 is recorded as the starting point of the first QRS wave, p1 is stored in the matrix QRS, and then step b is carried out to carry out the next step.

(2) Step b, detecting all elements between p 2 to py-1 in the matrix p one by one, detecting whether the element is the QRS wave starting point, marking p j as the j-1 point to be detected, if (p j-QRS < -1 >) is larger than or equal to 1.2 x Fs, then missing the QRS wave starting point between p j and QRS < -1 >. Then using new probability matrix Z [ WQRS [ -1] +1/5 × (Fs), … Wp [ j ] -1/5 × (Fs) ] to detect the starting point of QRS wave in the range, and detecting according to the fixed threshold method in step two, and at this time, the threshold value thr3 ═ thr1-0.1, and storing the detection result in the matrix QRS. If QRS is empty, lowering threshold to thr 4-thr 3-0.1 until QRS is not empty, if QRS is not empty, saving QRS and going to step c; if (pj-QRS < -1 >) < 1.2 xFs, pj is the starting point of QRS wave, store pj in matrix QRS, then go to step c to proceed to the next step.

(3) And c, if p < -1 > is less than or equal to N-Fs, detecting the QRS wave starting point between p < -1 > and N. Then the new probability matrix T [ WP [ -1] +1/5 × (Fs, … WN) ] is used to detect the starting point of QRS wave in the range, and the detection is performed according to the fixed threshold method in step two, and at this time, the threshold thr5 is thr1-0.1, and the detection result is stored in the QRS matrix. If QRS is empty, lowering threshold value thr 6-thr 5-0.1 until QRS is not empty, if QRS is not empty, storing QRS; if p-1 > N-Fs, p [ j ] is the starting point of QRS wave, store p [ j ] into the matrix QRS. Finally, QRS ═ QRS1, QRS2, …, QRSj, …, QRSz ], where z denotes the number of QRS wave starting points based on the dynamic threshold adaptive adjustment method; j denotes an index value. And extracting a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point according to the method of the step three.

In actual operation, the data to be detected is a multi-lead electrocardiosignal with T being 10s and the sampling frequency being Fs being 500Hz, and high-frequency and low-frequency noises are removed through discrete wavelet transformation based on db6 wavelet and a third-order Butterworth band-pass filter with 0.5-40 Hz. Taking a certain single lead signal X (T) ═ { x (T), T ═ 1,2, … N } of M lead electrocardiosignals, wherein N ═ T ═ Fs ═ 5000, then removing base line interference and high-frequency and low-frequency noise by a three-order Butterworth band-pass filter based on db6 wavelet discrete wavelet transform, intercepting each lead electrocardio data length as N, 80% of data as a training set, 20% of data as a test set, a detection point in the test set is a QRS wave starting point, a label is (1,0,0, 0), the QRS wave is a peak point, and a label is (0,1,0,0, 0); the QRS wave is the termination point, then markLabel is (0,0,1,0,0), the T wave is a peak point, label is (0,0,0,1,0), the T wave is an end point, label is (0,0,0,0,1), input data dimension is nx1, data label dimension is nx5, during which, if a certain sampling point is marked as a QRS wave start point, the sampling points of the first 0.075s and the last 0.075s of the sampling point are also considered as QRS wave start points, i.e. have the same label, and then the model loss function is:

where n is the number of batch samples per input to the network, 6 is the number of classes, and 5000 is the number of ECG sampling points. z represents the probability that the sampling time belongs to the class K, wherein K is {1,2,3,4,5,6}, K is equal to K, T is {0,1,2, …, 5000}, T is equal to T, and y is a sample label value, the model is trained by a Stochastic Gradient Descent (SGD) optimizer, and the model loss function is minimized through gradual iteration. Wherein the momentum, learning rate and batch size superparameters are 0.9, 0.005 and 64, respectively; when the loss function value is not reduced after the U-net-ECGCNN model and the U-net-ECGLSTM model are continuously trained for 20 rounds respectively, stopping training and respectively storing the model M and the model N at the moment, and when the model M and the model N need to be checked later, passing through

And

wherein P is_N1，P_N2，P_N3，P_N4，P_N5Representing the probability that the Nth sampling point based on the U-net-ECGCNN model respectively belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point; wherein Q_N1，Q_N2，Q_N3，Q_N4，Q_N5The probability that the Nth sampling point respectively belongs to a QRS wave starting point, a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point based on a U-net-ECGLSTM model is shown, a probability matrix V of which the sampling point at each moment is a corresponding characteristic point is obtained based on ensemble learning, and the expression is as follows:

wherein:

according to the feature matrix V obtained above, converting the feature matrix V into a matrix X consisting of 0 and 1, wherein the expression is as follows:

the process is as follows: if the probability value V_j1Greater than thr1, then X_j1Equal to 1, otherwise X_j1Equal to 0. For the detection of the starting point of the QRS wave, the probability matrix of each sampling point belonging to the starting point of the QRS wave is W1 ═ W₁,W₂,…W_j...W_N]The matrix consisting of 0 and 1 after conversion is X1 ═ X₁,X₂,…X_N]At this time, the specific step of finding the position of the starting point of the QRS wave is as follows: firstly finding the first position equal to 1 in X1, recording the position as X1, searching from the position of X1 until the position of the next X2 is found to be equal to 1, and the position of X2+1 is equal to 0, recording the position X2 at this time, wherein the middle position from X1 to X2 is the starting point p1 of the first QRS wave, then continuing to search backwards from the position X2+1, and recording the positions of all the starting points of the QRS waves in the same step, wherein the position p of the starting point of the QRS wave is [ p 1]₁,p₂,…,p_j,p_y]Wherein y represents the number of QRS wave starting points based on a fixed threshold method; j represents an index value; in a single lead signal x (t) of the above multi-lead electrocardiographic signal, the probability matrix of each sampling point belonging to the QRS wave starting point is W1 ═ based on the method of step two₁,W₂,…W_j...W_N]And applying the method in the second step to all the rest leads of the M-lead electrocardiosignals, wherein the total probability matrix of each sampling point of the M-lead electrocardiosignals belonging to the QRS wave initial point is as follows:

W_INexpressing the probability that the Nth sampling point on the I-th lead belongs to the QRS initial point, I is a lead index value, and a multi-lead mutual reference method based on electrophysiological knowledgeCalculating the probability mean value of the sampling points belonging to the QRS wave starting point on each lead, taking the probability mean value as the probability that the sampling points on the lead belong to the QRS starting point, and applying the probability mean value to N sampling points, wherein the probability matrix of each sampling point belonging to the QRS wave starting point is as follows: w ═ W₁,W₂,W₃,....,W_I,..W_N]Wherein

Wherein W_NRepresenting the probability that the Nth sampling point belongs to the QRS starting point; QRS wave starting point position p ═ p obtained by fixed threshold method₁,p₂,…,p_j,p_y]When p is₁≧ Fs, indicating that one or more QRS wave start points are missed, the new probability matrix Y ═ W is used₁,W₂,…W_p[0]+1/5*Fs]And detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr2 is equal to thr1-0.1, and storing the detection result in the matrix QRS. If QRS is empty, lowering threshold value thr 3-thr 2-0.1 until QRS is not empty, if QRS is not empty, storing QRS; when p is₁＜Fs，p₁Denote as the starting point of the first QRS wave, p₁Storing into matrix QRS, and detecting p 2 in matrix p one by one]To p [ y-1]All elements in the middle, detecting whether the element is the starting point of the QRS wave, and recording p [ j]Is the j-1 st point to be detected when (pj)]-QRS[-1]) Not less than 1.2 xFs, then pj]And QRS-1]In the process of missing detection of the QRS wave starting point, a new probability matrix Z is used as [ W ]_{QRS[-1]+1/5*Fs},…W_p[j]-1/5*Fs]And detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr3 is equal to thr1-0.1, and storing the detection result in the matrix QRS. If QRS is empty, the threshold is lowered to thr 4-thr 3-0.1 until QRS is not empty, if QRS is not empty, QRS is saved if (p [ j ] j)]-QRS[-1])＜1.2×Fs,p[j]For the starting point of QRS wave, p [ j ] is]Saving into the matrix QRS, then if p-1]Less than or equal to N-Fs, then p-1]And missing QRS wave starting point between N, using new probability matrix T ═ W_P[-1]+1/5*Fs,…W_N]Detecting the starting point of QRS wave in the range, and detecting according to the fixed threshold method in the above stepsAnd (4) measuring, wherein the threshold thr5 is thr1-0.1, and the detection result is stored in the QRS matrix. If QRS is empty, lowering threshold value thr 6-thr 5-0.1 until QRS is not empty, if QRS is not empty, storing QRS; if p is [ -1]]＞N-Fs,p[j]For the starting point of QRS wave, p [ j ] is]Storing the QRS matrix into a QRS matrix; last QRS ═ QRS₁,QRS₂,…,QRS_j,…,QRS_z]Wherein z represents the number of QRS wave starting points based on the dynamic threshold self-adaptive adjustment method; j denotes an index value. The same method can be used to extract the QRS wave peak point, QRS wave termination point, T wave peak point and T wave termination point.

And acquiring 12-lead electrocardiosignal sample data of a Robacco university electrocardio database (LUDB) and a Chinese Cardiovascular Disease Database (CCDD) (both public electrocardio databases) containing the position marking information of the characteristic points, and verifying according to the characteristic point extraction method. The block diagram of the method is shown in figure 1, the preprocessed 12-lead electrocardiosignals are input into trained CNN and LSTM models based on a U-net frame, and the positions of a QRS wave start point, a QRS wave peak point, a T wave peak point and a T wave end point are positioned by a dynamic threshold self-adaptive adjustment strategy and a multi-lead mutual reference method based on electrophysiological knowledge on the basis of a fixed threshold method. The results of feature point extraction on the two public data are shown in fig. 3 and 4, respectively; the effect of adding the dynamic threshold adaptive adjustment strategy compared with the multi-lead mutual-reference method is shown in 5. Furthermore, the LSTM network may be replaced by a network of bidirectional LSTM and gated round robin units (GRUs).

The invention has the beneficial effects that: according to the method for extracting the multi-lead electrocardiosignal characteristic points by combining deep learning and electrophysiological knowledge, morphological characteristics of electrocardiosignal heart beat and strong time sequence correlation characteristics of sampling moments can be represented through CNN and LSTM networks based on a U-net frame, and finer characteristics of each moment of a waveform can be enhanced through fusion of bottom layer information and high layer information, so that the purpose of providing precision is achieved, through setting of self-adaptive adjustment strategy correction based on dynamic threshold values, the threshold values are gradually reduced, the relevant characteristic points are searched again, the missing rate can be remarkably reduced, the purpose of reducing the missing rate is further achieved, through setting of multi-lead mutual reference method correction, the mean value of the probability that each sampling point of a plurality of lead sampling points belongs to the corresponding characteristic points is calculated, the accuracy rate of characteristic point positioning is further improved, and the noise interference is large in QRS wave group and ST-T section with complex shape, When a malignant arrhythmia event occurs, the characteristic points are accurately positioned, and the purposes of reducing misdiagnosis rate and missed diagnosis rate are achieved.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. A multi-lead electrocardiosignal characteristic point extraction method combining deep learning and electrophysiological knowledge is characterized by comprising the following steps: the method comprises the following steps:

(1) the multi-lead electrocardiosignal acquisition module:

acquiring a multi-lead electrocardiosignal through a signal acquisition sensor, an isolation circuit, an amplification circuit, a filter circuit and an analog/digital conversion circuit, and separating the multi-lead electrocardiosignal into the multi-lead electrocardiosignal, wherein the signal is recorded as X (t), the number of leads is M, the length is N, and the sampling frequency is Fs;

2. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 1, wherein the method comprises the following steps: the feature point extraction module establishes and trains a model after the data set is constructed and processed, and the establishment and training of the model adopts the following two steps: building a convolution neural network (U-net-ECGCNN) based on a U-net framework; and establishing a U-net framework-based long-short term memory network (U-net-ECGLSTM) and training the model.

3. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 2, wherein the method comprises the following steps: the U-net framework based convolutional neural network (U-net-ECGCNN) has 19 layers which comprise an encoding module, an underlying module, a decoding module and a Softmax layer; the coding module comprises 6 layers in total and consists of 6 one-dimensional convolution layers (1D-C), 3 batch normalization layers (BN), 3 Relu layers and three one-dimensional down-sampling layers (1D-P), wherein the length of a convolution kernel is S multiplied by 1, the numerical value of S in the first two layers is 31, then each two layers are reduced by 6, the number of the convolution kernels is 16 multiplied by 2K, the value of K is equal to 0 at first, the numerical values of every two convolution layers are increased by 1, and the down-sampling factor of each down-sampling layer is 2; the bottom layer module comprises 2 layers in total and consists of two 1D-C layers, 1 BN layer and a Relu layer; wherein the length and number of convolution kernels are 13 and 128 respectively; the decoding module consists of 10 one-dimensional convolutional layers, 6 BN layers, 6 Relu layers, three one-dimensional upsampling (1D-U) layers and three layer skipping link (SC) layers, the number of convolutional cores in the convolutional layers is 16 multiplied by 2K, the value of K is equal to 2 at first, the numerical value of each two convolutional layers is reduced by 1, and the upsampling factor of each upsampling layer is 2; the 1D-U can recover the detail information of each sampling point, the parameters of the detail information correspond to the coding module, and the SC can provide the extra information lost in the 1D-U stage; the Softmax layer outputs the probability that each sampling point belongs to the starting point, the peak point, the ending point of the QRS wave and the peak point and the ending point of the T wave respectively.

4. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 2, wherein the method comprises the following steps: the long-term and short-term memory network based on the U-net framework comprises an encoding module, a bottom layer module, a decoding module and a Softmax layer, and is different from a convolutional neural network based on the U-net framework in that the bottom layer module is replaced by two LSTM layers and used for extracting strong time sequence correlation characteristics of each sampling moment, and the number of hidden layer nodes in the two LSTM layers is 128.

5. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 1, wherein the method comprises the following steps: inputting an input signal { x (T), T1, 2, … N } into a U-net-ECGCNN model and a U-net-ECGLSTM model, and respectively obtaining probability matrixes P and Q of sampling points at each moment belonging to a QRS wave starting point, a QRS wave peak point, a QRS wave ending point, a T wave peak point and a T wave ending point, wherein the expressions are respectively as follows:

6. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 5, wherein the method comprises the following steps: acquiring a probability matrix V of which the sampling point at each moment is a corresponding characteristic point based on ensemble learning, wherein the expression is as follows:

wherein:

and converting it into a matrix consisting of 0 to 1, whose expression is:

if the probability value V_j1Greater than thr1, then X_j1Equal to 1, otherwise X_j1Equal to 0, for the detection of the starting point of the QRS wave, the probability matrix of each sample point belonging to the starting point of the QRS wave is W1 ═ W₁,W₂,…W_j...W_N]The matrix consisting of 0 and 1 after conversion is X1 ═ X₁,X₂,…X_N]The specific step of finding the position of the starting point of the QRS wave is as follows:

7. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 1, wherein the method comprises the following steps: the characteristic point correction module corrects the position of the characteristic point by a multi-lead mutual reference method based on electrophysiological knowledge and adaptively adjusts a strategy based on a dynamic threshold valueThe position of the positive characteristic point, wherein a certain single lead signal X (t) of the 12-lead electrocardiosignals in the step (1), and the probability matrix of each sampling point belonging to the starting point of the QRS wave based on the method in the step (2) is W1 ═ W₁,W₂,…W_j...W_N]And (3) applying the method in the step (2) to all the rest leads of the M-lead electrocardiosignals, wherein the total probability matrix of each sampling point of the M-lead electrocardiosignals belonging to the QRS wave initial point is as follows:

W＝[W₁,W₂,W₃,....,W_I,..W_N]

wherein

8. The method for extracting the characteristic points of the multi-lead electrocardiosignals combining deep learning and electrophysiological knowledge as claimed in claim 7, wherein the method comprises the following steps: based on the position of the characteristic point corrected by the self-adaptive adjustment strategy of the dynamic threshold, aiming at the position p of the starting point of the QRS wave, which is p ═ p₁,p₂,…,p_j,…p_y]Detecting all elements in the matrix p one by setting different threshold values, detecting whether the element is a QRS wave starting point, and storing the result in the matrix QRS, wherein the method comprises the following specific steps:

step a, if p [1]]≧ Fs, indicating that one or more QRS wave start points are missed, the new probability matrix Y ═ W is used₁,W₂,…W_p[0]+1/5*Fs]Detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr2 is thr1-0.1, the detection result is stored in the matrix QRS, if the QRS is empty, the threshold value thr3 is thr2-0.1, until the QRS is not empty, if the QRS is not empty, the QRS is stored and the step b is carried out; if p [1]]＜Fs，p[1]Marking as the starting point of the first QRS wave, and p1]Storing the QRS matrix into a QRS matrix, and then turning to the step b to carry out the next step;

step b, detecting p 2 in matrix p one by one]To p [ y-1]All elements in the middle, detecting whether the element is the starting point of the QRS wave, and recording p [ j]Is the j-1 st point to be detected, if (pj)]-QRS[-1]) Not less than 1.2 xFs, then pj]And QRS-1]In the process of missing detection of the QRS wave starting point, a new probability matrix Z is used as [ W ]_{QRS[-1]+1/5*Fs},…W_p[j]-1/5*Fs]Detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr3 is thr1-0.1, the detection result is stored in the matrix QRS, if the QRS is empty, the threshold value thr4 is thr3-0.1, until the QRS is not empty, if the QRS is not empty, the QRS is stored and the step c is carried out; if (p [ j)]-QRS[-1])＜1.2×Fs,p[j]For the starting point of QRS wave, p [ j ] is]Storing the QRS matrix into a QRS matrix, and then turning to the step c to carry out the next step;

step c, if p [ -1]]Less than or equal to N-Fs, then p-1]And missing QRS wave starting point between N, using new probability matrix T ═ W_P[-1]+1/5*Fs,…W_N]Detecting the starting point of the QRS wave in the range, detecting according to the fixed threshold method in the step two, wherein the threshold value thr5 is equal to thr1-0.1, the detection result is stored in the QRS matrix, if the QRS is empty, the threshold value thr6 is decreased to be equal to thr5-0.1 until the QRS is not empty, and if the QRS is not empty, the QRS is stored; if p is [ -1]]＞N-Fs,p[j]For the starting point of QRS wave, p [ j ] is]Storing into matrix QRS, and storing the QRS ═ QRS₁,QRS₂,…,QRS_j,…,QRS_z]Wherein z represents the number of QRS wave starting points based on the dynamic threshold self-adaptive adjustment method; j represents an index valueAnd extracting a QRS wave peak point, a QRS wave termination point, a T wave peak point and a T wave termination point according to the same method.