CN111956215A - QRS point detection method of low-quality electrocardiogram - Google Patents

Abstract

The invention discloses a QRS point detection method of electrocardiosignals, which comprises the following steps: 1) denoising an original electrocardiosignal; 2) obtaining candidate QRS points; 3) removing non-QRS points in the candidate QRS points and generating a data set; 4) screening QRS point coordinates through a convolutional neural network; performing bottom convolution on the electrocardiosignal by using a modern convolutional neural network algorithm, and further extracting features from the features obtained by the bottom convolution of each lead through an LSTM layer and a Dense layer to obtain final features which are input into a classifier for classification to obtain a QRS point classification result; the method obtains higher accuracy rate for identifying the QRS point of the electrocardiosignal.

Description

A method for detecting QRS points of low-quality electrocardiogram

technical field

The invention relates to the technical field of medical signal processing, more specifically a QRS point detection method for low-quality electrocardiogram.

Background technique

With the continuous improvement of the quality of life in my country, people pay more and more attention to health, especially the prevention of cardiovascular diseases. The electrocardiogram is an important means of detecting cardiovascular disease. In the process of using electrocardiogram to detect cardiovascular diseases, doctors will first determine each waveform in the electrocardiogram, and the QRS complex, as the most obvious feature of the electrocardiogram, often plays a crucial role in the determination of each waveform of the electrocardiogram. However, with the pursuit of more convenient and fast ECG detection, wearable ECG monitoring equipment has emerged. Although it allows people to collect ECG more conveniently, the quality of the ECG collected is caused by the movement of the person being collected and other environmental factors. On the low side, the difficulty of detecting the QRS complex is greatly increased.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the problem of low quality of electrocardiogram caused by problems of the environment and the acquisition equipment itself, and thus the problem of low recognition accuracy of QRS points of electrocardiogram, and to provide a method for detecting QRS points of low quality electrocardiogram.

ECG signal QRS point detection method, which includes:

1) Denoising of the original ECG signal

型带通滤波，得到去噪信号矩阵x；读取每一个原始心电信号的人工标记的QRS点矩阵ref，为下文制作标签做准备Read the original ECG signal e, and perform Chebyshev on the original ECG signal e

Type band-pass filtering to obtain the denoised signal matrix x; read the manually marked QRS point matrix ref of each original ECG signal to prepare for the labeling below

2) Get candidate QRS points

Hilbert transform is performed on the denoised signal matrix x to obtain the transformed signal matrix x1, and the absolute value of the transformed signal matrix x1 is obtained to obtain the envelope signal matrix x2 of the denoised signal. Identify the peak points and trough points of the envelope signal matrix x2, and store the set of peak points and trough points of the envelope signal matrix x2 into the matrix H_place1 as the envelope amplitude value of the over-detected QRS points of the ECG signal, and store the The subscript set of the peak points and trough points of the envelope signal matrix x2 is stored in the matrix R_place1 as the coordinate points of the over-detected QRS points of the ECG signal. The elements of matrix R_place1 and matrix H_place1 are in one-to-one correspondence.

3) Remove non-QRS points in candidate QRS points and generate datasets

a. Remove the non-QRS points detected by the refractory period in the matrices R_place1 and H_place1. The non-QRS point detected by the refractory period here refers to the difference between the two adjacent elements R1 and R2 of the matrix R_place1. The value range of the period is any integer between 0 and 300, and the two adjacent elements H1 and H2 in the matrix H_place1 have smaller values, and then the matrices R_place3 and H_place3 are obtained. The subscript of the element R1 in the matrix R_place1 is the same as the subscript of the element H1 in the matrix H_place1, and the subscript of the element R2 in the matrix R_place1 is the same as the subscript of the element H2 in the matrix H_place1.

b. In the original ECG signal e, take the elements in R_place3 as the subscript, and intercept P points forward. When intercepting forward, the elements of the subscript need to be included, and Q points are intercepted backward. The heart beat intercepts the candidate QRS point raw data sample T1 of W=P+Q points,

c. In the envelope signal matrix x2, take the elements in R_place3 as the subscript, and intercept P points forward. When intercepting forward, the elements of the subscript need to be included, and Q points are intercepted backward. The heart beat intercepts the candidate QRS point envelope data samples T2 of W=P+Q points,

d. Perform one-dimensional median filtering on the obtained candidate QRS point raw data sample T1 to obtain the candidate QRS point filtered signal T3,

e. Perform horizontal splicing of T1, T2, and T3 to obtain a (1*C)-dimensional candidate QRS point data sample T, where C is equal to 3*W, and then T is used as the input X of the convolutional neural network model. All ECG signals are intercepted by the above interception method to all elements in R_place3 to form a data set U, wherein each sample in the data set U is the candidate QRS point data after preliminary screening of the above 1*(C) dimension sample.

f. The method for determining the label of each intercepted candidate QRS point data sample is to compare the elements in R_place3 corresponding to the candidate QRS point data sample with the read QRS point matrix ref of the original ECG signal in turn , if the difference between the elements in R_place3 and each element in the matrix ref is less than 37, the label data label of the intercepted over-detection class QRS point data sample is judged to be 1, otherwise it is 0.

4) Screening QRS point coordinates through convolutional neural network

a. Build a convolutional neural network model

The core of the convolutional neural network is mainly composed of three bottom convolutional layers in series, one LSTM layer, one Attention layer, and four Dense layers.

b. Training convolutional neural network parameters

Initialize the parameters of the convolutional neural network, randomly select 80% of the samples from the sampled data set U as the training set, and 20% of the data set U as the test set; input the ECG signal samples in the training set into the In the initialized neural network, iterates with the goal of minimizing the cost function to generate and save the parameters of the convolutional neural network; when extracting the training set and the test set, the elements of the label are also allocated to the corresponding matrix respectively, where the labels of the test set are put into the matrix y_label.

c. Automatic identification of test set samples

Input the divided test set samples into the convolutional neural network and run it to obtain the 2-dimensional prediction value vector output y_pred corresponding to the test set samples, and then select the index value of the maximum value of the 2-dimensional prediction vector output as whether the point is a QRS point. Prediction category, if the index is 1, the corresponding over-detection class QRS point data sample is the detected QRS point data sample, and the corresponding element in R_place3 is the detected QRS point.

The parameters of the convolutional neural network are: input X is the ECG signal sample, each candidate QRS point data sample is 1*W dimension, W is the number of points based on each candidate QRS point data sample; Input to the bottom convolutional layer, where each bottom convolutional layer contains a convolutional layer, the output of each convolutional layer unit is connected in series with a BN layer, an excitation unit operation and a pooling layer operation; the first The number of convolution kernels of each convolutional layer unit is 64, the size of the convolution kernel is 3, the convolution method is the same, the excitation unit after the BN layer unit is the elu function, and the pooling kernel size of the pooling layer unit is 2. The pooling step size is 2; the dimension of the feature map after the first layer of pooling units is (C/2)*64; the number of convolution kernels of the second convolutional layer unit is 256, and the size of the convolution kernel is 3 , the convolution method is the same, the excitation unit after the BN layer unit is the elu function, the pooling kernel size of the pooling layer unit is 2, and the pooling step size is 2; the dimension of the feature map after the second layer pooling unit is (C4)*256, the number of convolution kernels of the third convolution layer unit is 32, the size of the convolution kernel is 3, the convolution method is the same, the excitation unit after the BN layer unit is the elu function, and the pooling layer unit The size of the pooling kernel is 2, and the pooling step size is 2; the dimension of the feature map after the third layer of pooling units is (C/8)*256, and then through an LSTM layer, the output of the LSTM layer is defined as 32, The dimension of the feature map after passing through the LSTM layer is (C/8)*32, and then through an Attention layer, the dimension of the feature map obtained is 1*32, and after two Dense layers, the dimension of the feature obtained in turn is 1*16, 1*4, the output y_pred is finally obtained through a Dense layer, and the final classification result is obtained by processing y_pred and is stored in the matrix RE;

The iterations are: iteratively update the training parameters once, until the loss value and the accuracy rate of the convolutional neural network are stabilized around a certain value, stop training and save the training parameters and model structure information of the current network.

The present invention provides a method for detecting QRS points of an ECG signal, which includes: 1) denoising the original ECG signal; 2) obtaining candidate QRS points; 3) removing non-QRS points in the candidate QRS points and generating a data set; 4) The QRS point coordinates are screened through the convolutional neural network; the underlying convolution of the ECG signal is performed using the modern convolutional neural network algorithm, and then the features obtained by the underlying convolution of each lead are further extracted through the LSTM layer and the Dense layer. The final feature is obtained and then input to the classifier for classification to obtain the QRS point classification result. The recognition of QRS points of ECG signal obtained by this method has a high accuracy. Its confusion matrix is as follows:

Description of drawings

Figure 1 Schematic diagram of the bottom convolutional layer; where X is the input, c is the convolutional layer, b is the BN layer, p is the maxpling layer, l is the lstm layer, a is the excitation unit, d is the dence layer, Attention is the Attention layer, y_pred for output.

Detailed ways

Embodiment 1 QRS point detection method of ECG signal

The present invention will be further described below in conjunction with specific embodiments.

A specific example is the training dataset released at the 2nd China Physiological Signal Challenge in 2019, which collected 2,000 single-lead ECG recordings from cardiovascular disease (CVD) patients, each lasting 10 s. The tester contains similar ECG recordings of the same length, which are not public and will remain private for scoring during the challenge and for a period of time thereafter. ECG recordings were obtained from multiple sources using a variety of instruments, although in all cases they are represented here at a 500 Hz sampling rate. All records are provided in MATLAB format (each record includes two .mat files: one with ECG data and the other with the corresponding QRS annotation file). This example is implemented through the software system (windows linux) working on the computer and the well-known Matlab simulation environment and python simulation environment in the industry.

The detailed steps of this embodiment are as follows:

1. Denoising of the original ECG signal

(1) Variable setting and data preparation

Create a new tuple ref with a size of 1*2000, and use the load function to load the 2000 pieces of ECG data from the training data set released in the 2nd China Physiological Signal Challenge in 2019 into the environment variables, and load the ECG signals of each person. The latter variable is a matrix e with a size of 1*5000, and each original ECG signal matrix e is read in turn to form a variable ecg of 2000*5000, where each row is the first channel of each ECG data in the database. ECG signal. At the same time, use the load() function to read the manually marked QRS points of each original ECG signal and store them in the tuple ref{ii}, where ii is 1 to 2000, corresponding to the manually marked QRS points of 2000 original ECG signals respectively. QRS points.

型滤波器参数准备(2) Chebyshev

Type filter parameter preparation

型带通滤波器滤波后的信号。设滤波器阶数n为2，阻带衰减Rs为30Hz，分辨率fs为500Hz。通带频率为5Hz到50Hz。则通带频率矩阵W为由数字5和数字50组成的1*2的矩阵。将参数n,Rs,W输入到MATLAB内置函数cheby2()中，得到切比雪夫二型滤波器参数a,b。然后将参数a,b和待滤波信号ecg(i,:)一起输入到MATLAB内置函数flitflit()中，输出为与输入的一维信号相同的矩阵。最后将数据库的每一个心电数据的第一通道心电信号进行滤波后的信号依次存放到变量y中。Create a new 2000*5000 variable y for storing Chebyshev

The signal after filtering by the type band-pass filter. The filter order n is set to 2, the stopband attenuation Rs is 30Hz, and the resolution fs is 500Hz. The passband frequency is 5Hz to 50Hz. Then the passband frequency matrix W is a 1*2 matrix composed of numbers 5 and 50. Input the parameters n, Rs, W into the MATLAB built-in function cheby2() to get the Chebyshev type 2 filter parameters a, b. Then the parameters a, b and the signal to be filtered ecg(i,:) are input into the MATLAB built-in function flitflit(), and the output is the same matrix as the input one-dimensional signal. Finally, the filtered signal of the first channel ECG signal of each ECG data in the database is sequentially stored in the variable y.

The specific operations are as follows:

when i takes 1 to 44

[b,a] = cheby2(N,Rs,W*2/fs);

y(i,:)= filtfilt(b,a,ecg(i,:));

2. Generate QRS point data samples

(1) Variable setting and data preparation

Create a new temporary storage denoising signal matrix b with a size of 1*5000, which is used to temporarily store the filtered signal intercepted when traversing the matrix y in step (2). A newly created tuple R_place1 with a size of 1*2000 is used to store the detected over-detection class QRS point coordinates. Each element of the tuple R_place1 is an over-detected QRS point coordinate point matrix detected from the ECG signal of the first channel of an ECG data in the database. A new tuple H_place1 with a size of 1*2000 is used to store the detected over-detection class QRS point amplitude values, where each element of the tuple H_place1 is the first channel ECG signal detection of an ECG data from the database The resulting over-detection class QRS point amplitude value matrix.

Step (2) is performed when i=1, 2, 3...44:

(2) Temporarily store the denoising signal matrix, let b=y(i,:), use the built-in function hilbert() of MATLAB to calculate the Hilbert transform of the temporarily stored denoising signal matrix b, and obtain the transformed signal x1. The absolute value of the signal x1 is taken to obtain the envelope signal x2 of the denoised signal. And store the envelope signal in the matrix y_ah(i,:). Use the MATLAB built-in function fingpeaks() to identify the peak point coordinates of the envelope signal x2 as the over-detection QRS point coordinate point set and the amplitude value set of the ECG signal. The former is stored in R_place1{i}, and the latter is stored in H_place1 in {i}.

The specific operations are as follows:

b = y(i,:);

x1= hilbert(b);

x2 = abs(x1);

[H_place1{i},R_place1{i}]= fingpraks(x2);

y_ah(i,:) = x2;

3. Remove non-QRS points in candidate QRS points and generate datasets

Steps (1) (2) (3) (4) are executed when i=1, 2, 3...2000:

(1) Data preparation

The new matrix R_place is equal to R_place1{i}, the new matrix R_place2 is equal to R_place1{i}, the new matrix e is equal to y_ah(i,:), the artificially set refractory period X is set to 40, and the new matrix H_place is equal to H_place1{i}. A new empty matrix pr is used to store non-QRS points.

(2) Refractory period to remove non-QRS points

Use the MATLAB built-in function min() to find the minimum value p of H_place, and return the subscript pm of the minimum value p in the matrix H_place1, traverse the array R_place2, find the same element as R_place (pm) and return the element subscript pm1, using The MATLAB built-in function diff() calculates the difference between the element R_place2(pm1) and its adjacent elements, and returns the matrix dif. If all elements of the matrix dif are smaller than the artificially set refractory period X, the element R_place2 (pm1) is a non-R point element and is deleted. Then delete the elements with the subscript pm in the matrices R_place1 and H_place1 to ensure that the minimum value can be smoothly updated when the matrix H_place1 is looking for the minimum value and the elements of the matrices R_place1 and H_place1 are consistent. Repeat the above steps in sequence until the difference between every two adjacent elements of R_place1{i} is less than B.

(3) Data set generation

a. In the original ECG signal ecg(i,:), the elements in R_place1{i} are used as subscripts, and 149 points are intercepted forward. After 150 points are intercepted, each element in R_place1{i} intercepts 300 points of candidate QRS point raw data samples T1.

b. In the envelope signal x2(i,:), the elements in R_place1{i} are used as subscripts, and 149 points are intercepted forward. 150 points are intercepted, and each element in R_place1{i} intercepts 300 points of candidate QRS point envelope data samples T2.

c. The QRS point data sample T1 is subjected to 30-order median filtering through the one-dimensional median filter function medfilt1() of matlab to obtain the candidate QRS point filtered data sample T3.

d. Perform horizontal splicing of the original data sample T1 of the candidate QRS point, the envelope data sample T2 of the candidate QRS point, and the filtered data sample T3 of the candidate QRS point to obtain a QRS mixed data sample with a dimension of 1*900

Then each element in R_place1{i} gets 1*900 points as a sample of each candidate QRS point data sample T. Then do the same for all the elements in R_place1{i}. Get a dataset containing (1*900)*114926-dimensional data, because each sample is (1*900)-dimensional, and since each sample is intercepted according to the element position in R_place1, 114926 is used for interception The number of elements in R_place1{i}, that is, the number of samples. Each sample is a 1*900 candidate QRS point data sample T, which is used as the input of the convolutional neural network.

e. The label of each QRS point data sample T is obtained by sequentially reading the elements in R_place1{i} and subtracting all elements in ref{i} to obtain matrix l_dif, and then go to the absolute value of matrix l_dif to obtain matrix l_ad , If there are elements less than 37 in the matrix l_ad, judge that the label of the QRS point data matrix corresponding to the elements in R_place1{i} is 1, otherwise it is 0,

4. Screening QRS point coordinates through convolutional neural network

(1) Build a convolutional neural network model

The input of the convolutional neural network model is a QRS point mixed data sample T, where T is a (1*900) dimensional heart candidate QRS point data sample output by the preprocessing part, and 900 is the intercepted candidate QRS point data sample number of points. . Input a 1*900 dimension of the input ECG signal sample, that is, the candidate QRS point data sample into the bottom convolution layer, that is, the candidate QRS point data sample enters the bottom convolution layer for processing, wherein each bottom convolution layer contains a convolution layer. layer, the output of each convolution layer unit is connected in series with a BN layer, an excitation unit operation and a pooling layer operation; the number of convolution kernels of the first convolution layer unit is 64, and the size of the convolution kernel is 3. The convolution method is the same, the excitation unit after the BN layer unit is the elu function, the pooling kernel size of the pooling layer unit is 2, and the pooling step size is 2; the dimension of the feature map after the first layer of pooling unit is (900/2)*64; the number of convolution kernels of the second convolution layer unit is 256, the size of the convolution kernel is 3, the convolution method is the same, the excitation unit after the BN layer unit is the elu function, the pool The pooling kernel size of the layer unit is 2, and the pooling step size is 2; the dimension of the feature map after the second layer pooling unit is (900/4)*256, and the convolution kernel of the third convolution layer unit The number is 32, the size of the convolution kernel is 3, the convolution method is the same, the excitation unit after the BN layer unit is the elu function, the size of the pooling kernel of the pooling layer unit is 2, and the pooling step size is 2; The dimension of the feature map after the three-layer pooling unit is (900/8)*256, and then through an LSTM layer, the output of the LSTM layer is defined as 32, and the dimension of the feature map after the LSTM layer is (900/8)*32, Then through an attention layer, the dimension of the feature map is 1*32, and finally through four Dense layers, the output y_perd is obtained; the model is built using the tensorflow open source framework and python language.

The network parameters of the bottom convolutional layer can be viewed in Table 2.

The neural network is built using the functional model in the keras framework, that is, the Model function is imported from the keras.models module, the input of the Model is set to the above-mentioned single-lead ECG signal sample X, and the output is a prediction vector y_pred with dimension 2 . The one-dimensional convolution layer is constructed by importing the tensorflow.keras.layers.Conv1D function, and the one-dimensional pooling layer is constructed by importing the tensorflow.keras.layers.MaxPool1D function. The BatchNormalization layer is tensorflow.keras.layers.BatchNormalization, and the fully connected layer is BatchNormalization .Dense.

(2) Parameters for training the convolutional neural network model

First, the training parameters of the neural network model are initialized, and the sampled signal is divided into training set samples and test set samples, and the divided data set U is shown in Table 3. The single-lead ECG signal sampled in the training set is input into the initialized convolutional neural network model, and the root-mean-square logarithmic error is used as the cost function in the convolutional neural network. The mean_squared_logarithmic_error function is used in Tensorflow.Keras, an object model is instantiated through the constructed functional model Model in the neural network, and the parameter loss is set as 'mean_squared_logarithmic_error' in the model.compile function. And use the Adam optimizer to iterate with the goal of minimizing the cost function, and optimize by setting the parameter optimizer to 'Adam' in the model.compile function to generate the deep neural network and save it as a file my_model.hd5 with a hd5 suffix; Wherein, the training parameters are updated once every iteration. Until the loss value and the accuracy rate of the deep neural network described at the end are stabilized around a certain value, the training can be stopped and the training parameters and model structure information of the current network can be saved. The neural network was trained for a total of 1000 batches of 64 samples each.

(3) Automatic identification of test set samples

Input all the divided test set samples into the saved convolutional neural network model1.hd5, and run the convolutional neural network to obtain the 4-dimensional predicted value vector output y_pred corresponding to the test set samples, through python The argmax function in the built-in package numpy inputs the y_pred matrix into the function, and finally obtains the subscript of the maximum value of each four-dimensional predicted value vector. matrix RE. Then count the same number of samples n of the classification result RE as y_label, and divide n by the total number of samples in the test set to obtain the final accuracy of the convolutional neural network. The over-detection class QRS point data sample corresponding to the category with the label of 1 in RE is the detected QRS point data sample, and the corresponding element in R_place1 is the detected QRS point.

Claims (3)

1. ECG signal QRS point detection method, which includes: 1) Denoising of the original ECG signal Read the original ECG signal e, and perform Chebyshev on the original ECG signal e

Type band-pass filtering to obtain the denoised signal matrix x; read the manually marked QRS point matrix ref of each original ECG signal to prepare for the labeling below 2) Get candidate QRS points Perform Hilbert transform on the denoised signal matrix x to obtain the transformed signal matrix x1, take the absolute value of the transformed signal matrix x1 to obtain the envelope signal matrix x2 of the denoised signal; identify the peak points of the envelope signal matrix x2 and the The trough point, and the set of peak points and trough points of the envelope signal matrix x2 are stored in the matrix H_place1 as the envelope amplitude value of the over-detected QRS point of the ECG signal, and the peak points and the peak points of the envelope signal matrix x2 and The subscript set of the trough point is used as the coordinate point of the over-detection QRS point of the ECG signal, and is stored in the matrix R_place1; the elements of the matrix R_place1 and the matrix H_place1 are in one-to-one correspondence; 3) Remove non-QRS points in candidate QRS points and generate datasets a. Remove the non-QRS points detected by the refractory period in the matrices R_place1 and H_place1. The non-QRS point detected by the refractory period here refers to the difference between the two adjacent elements R1 and R2 of the matrix R_place1. The value range is any integer between 0 and 300, and the two adjacent elements H1 and H2 in the matrix H_place1 have smaller values, and then the matrices R_place3 and H_place3 are obtained; where the element R1 is in the subscript of the matrix R_place1 and The subscript of element H1 in matrix H_place1 is the same, and the subscript of element R2 in matrix R_place1 is the same as that of element H2 in matrix H_place1; b. In the original ECG signal e, take the element in R_place3 as the subscript, and intercept P points forward. When intercepting forward, the elements of the subscript need to be included, and Q points are intercepted backward. The heart beat intercepts the candidate QRS point raw data sample T1 of W=P+Q points, c. In the envelope signal matrix x2, take the elements in R_place3 as the subscript, and intercept P points forward. When intercepting forward, the elements of the subscript need to be included, and Q points are intercepted backward. The heart beat intercepts the candidate QRS point envelope data samples T2 of W=P+Q points, d. Perform one-dimensional median filtering on the obtained candidate QRS point raw data sample T1 to obtain the candidate QRS point filtered signal T3, e. Horizontally splicing T1, T2, and T3 to obtain a (1*C)-dimensional candidate QRS point data sample T, where C is equal to 3*W, and then T is used as the input X of the convolutional neural network model; The above-mentioned interception method intercepts all elements in R_place3 to form a data set U, wherein each sample in the data set U is a candidate QRS point data sample after the preliminary screening of the above-mentioned 1*(C) dimension; f. The method for determining the label of each intercepted candidate QRS point data sample is to compare the elements in R_place3 corresponding to the candidate QRS point data sample with the read QRS point matrix ref of the original ECG signal in turn , if the difference between the elements in R_place3 and each element in the matrix ref is less than 37, it is judged that the label data label of the intercepted over-detection class QRS point data sample is 1, otherwise it is 0; 4) Screening QRS point coordinates through convolutional neural network a. Build a convolutional neural network model The core of the convolutional neural network is mainly composed of three bottom convolutional layers in series, one LSTM layer, one Attention layer, and four Dense layers; b. Training convolutional neural network parameters Initialize the parameters of the convolutional neural network, randomly select 80% of the samples from the sampled data set U as the training set, and 20% of the data set U as the test set; input the ECG signal samples in the training set into the In the initialized neural network, iterates with the goal of minimizing the cost function to generate and save the parameters of the convolutional neural network; when extracting the training set and the test set, the elements of the label are also allocated to the corresponding matrix respectively, The label of the test set is put into the matrix y_label; c. Automatic identification of test set samples Input the divided test set samples into the convolutional neural network and run it to obtain the 2-dimensional prediction value vector output y_pred corresponding to the test set samples, and then select the index value of the maximum value of the 2-dimensional prediction vector output as whether the point is a QRS point. Prediction category, if the index is 1, the corresponding over-detection class QRS point data sample is the detected QRS point data sample, and the corresponding element in R_place3 is the detected QRS point. 2. ECG signal QRS point detection method according to claim 1, is characterized in that: described convolutional neural network parameter is: input X is ECG signal sample, and each candidate QRS point data sample is 1*W dimension, W is the number of points in each candidate QRS point data sample; the candidate QRS point data samples are input into the underlying convolutional layer, where each underlying convolutional layer contains a convolutional layer, and the output of each convolutional layer unit A BN layer, an excitation unit operation and a pooling layer operation are connected in series at the ends; the number of convolution kernels of the first convolution layer unit is 64, the size of the convolution kernel is 3, and the convolution method is the same, BN layer The excitation unit after the unit is the elu function, the pooling kernel size of the pooling layer unit is 2, and the pooling step size is 2; the dimension of the feature map after the first layer of pooling units is (C/2)*64; The number of convolution kernels of the two convolutional layer units is 256, the size of the convolution kernel is 3, the convolution method is the same, the excitation unit after the BN layer unit is the elu function, and the pooling kernel size of the pooling layer unit is 2 , the pooling step size is 2; the dimension of the feature map after the second layer pooling unit is (C4)*256, the number of convolution kernels of the third convolution layer unit is 32, and the size of the convolution kernel is 3, The convolution method is the same, the excitation unit after the BN layer unit is the elu function, the pooling kernel size of the pooling layer unit is 2, and the pooling step size is 2; the dimension of the feature map after the third layer pooling unit is ( C/8)*256, and then pass through an LSTM layer, the output of the LSTM layer is defined as 32, the dimension of the feature map after passing through the LSTM layer is (C/8)*32, and then through an Attention layer, the resulting feature map dimension It is 1*32, after two Dense layers, the feature dimensions obtained in turn are 1*16, 1*4, and finally the output y_pred obtained through a Dense layer, the final classification result obtained by processing y_pred is stored in the matrix RE . 3. The electrocardiographic signal QRS point detection method according to claim 2, wherein the iteration is: iteratively update a training parameter once, until the loss value and the accuracy rate of the final convolutional neural network are stable at a certain Near the value, stop training and save the training parameters and model structure information of the current network.

Images