A ventricular premature beat localization method integrating deep learning of time domain and frequency domain features

Abstract

The present invention discloses a method for locating premature ventricular beats by integrating deep learning of time domain and frequency domain features, and uses the patient's 12-lead body surface potential data to determine the position of the ectopic pacemaker of the heart's premature ventricular beats. The present invention introduces the idea of ​​deep learning into the positioning of ectopic pacemakers, and by learning a large amount of clinical data, allows the computer model to find the hidden relationship between the electrocardiogram signal and the position of the ectopic pacemaker, replacing the rough manual judgment to achieve a more accurate effect. The present invention combines several types of models in the field of natural language processing, creatively integrating time domain and frequency domain features, and provides a more accurate method for deep learning models to process time series signals for regression prediction. The present invention realizes the positioning prediction of the origin point of premature ventricular beats from a data-driven perspective, and provides a non-invasive, safe, time-saving and labor-saving method for clinical catheter ablation surgery.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrocardiogram information analysis, and specifically relates to a method for locating premature ventricular beats by integrating deep learning of time domain and frequency domain features. Background Art

[0002] According to the World Health Organization, cardiovascular disease is the world's leading cause of death and a major threat to human health. Among the many cardiovascular diseases, arrhythmia is one of the most common symptoms. Clinically, ventricular premature beats, as a common arrhythmia, are increasingly receiving attention from doctors and people. Ventricular premature beats (abbreviated as VP) refer to ventricular electrical activity that occurs earlier than the basic heart rhythm (mostly sinus rhythm), which is caused by the His bundle bifurcation and the following ectopic pacemaker, and can trigger severe ventricular fibrillation and ventricular tachycardia and even sudden cardiac death. Therefore, it is particularly important to diagnose the ventricular tachycardia pacemaker in time and perform ablation surgery and other treatments on the patient accordingly.

[0003] Clinically, electrocardiograms are often used as the main basis for diagnosing heart diseases. Electrocardiograms are a technology that uses an electrocardiograph to record the changes in electrical activity generated by each cardiac cycle from the body surface. We can think of the heart as a discharge tissue. During the excitement of the heart, myocardial cells generate action potentials and then generate electrical signals. The conductive tissues or body fluids around the heart can transmit the electrical signals to the body surface, which are then received by external electrodes on the body surface and become electrocardiograms. Therefore, electrocardiograms can reflect the electrophysiological activities of the heart. The 12-lead electrocardiograms of ventricular premature beats from different origins have different characteristics, so they can provide guidance for doctors to determine the location of ventricular premature beats.

[0004] Although the current clinical research on ventricular premature beat positioning based on 12-lead electrocardiograms is relatively mature, there are still certain limitations in manual judgment, that is, the positioning accuracy of ventricular tachycardia is not high, and only a rough range of ventricular premature beats can be judged - such as the left and right ventricular outflow tracts, mitral valve rings, tricuspid valve rings, etc., while clinical ablation surgery requires more accurate location information. Currently, the methods that can provide more accurate positioning of abnormal pacemakers are mostly invasive means - usually requiring the placement of probes on the surface of the heart to determine the potential distribution of the heart, but this method requires a relatively large amount of time and money on the one hand, and on the other hand, it poses a certain threat to the patient's health.

[0005] By investigating the technologies related to the location of abnormal excitation points in the heart, it is found that the existing technologies are roughly divided into two categories: one is to use data-driven to achieve non-invasive abnormal excitation point location, such as a method based on autoencoder mentioned in the Chinese patent with publication number CN107007279A, which requires manual selection of QRS wave groups and input into the stacked autoencoder, and obtains the predicted (x, y, z) coordinates after training; from the perspective of the completeness of the input features, this method does not only use the QRS wave group of the ECG signal, but the information of the P wave, T wave, etc. is discarded, but they also contain information about ectopic pacemakers, which may reduce the accuracy of the prediction; from the perspective of algorithm selection, the ECG signal is a time series signal, and the signal has a long-distance dependency relationship. The algorithm selected in the field of natural language processing can extract more non-regional information between ECG signals than the patented technology. The second type is to establish a personalized heart-trunk model, solve the heart surface potential map through 64-lead ECG signal inversion, and then infer the location of the abnormal excitation point. For example, Chinese patent publication number CN105796094A discloses a method for locating abnormal excitation points of ventricular premature beats based on ECGI. The disadvantage of this type of method is that the procedure is complicated. First, the 64-lead ECG is not used as frequently as the 12-lead ECG in clinical practice. Secondly, it is necessary to specially collect three-dimensional CT or MRI images of the patient's heart and trunk. Summary of the invention

[0006] In view of the above, the present invention provides a ventricular premature beat localization method that integrates deep learning of time domain and frequency domain features, which can use 12-lead body surface potential data to well locate the three-dimensional coordinates of the ventricular premature beat pacemaker and meet the accuracy requirements of clinical surgery.

[0007] A ventricular premature beat location method integrating time domain and frequency domain feature deep learning comprises the following steps:

[0008] (1) Collect 12-lead surface electrocardiogram data of a large number of patients with different ventricular premature beats, and mark the three-dimensional coordinates of the cardiac ectopic pacemaker corresponding to each set of electrocardiogram data;

[0009] (2) preprocessing the 12-lead surface electrocardiogram data to obtain time domain information of the electrocardiogram signal;

[0010] (3) performing a one-dimensional Fourier transform on the 12-lead surface electrocardiogram data lead by lead to obtain frequency domain information of the electrocardiogram signal including amplitude and phase;

[0011] (4) combining the time domain information and frequency domain information of the ECG signal as a set of samples, and dividing all samples into a training set, a validation set, and a test set;

[0012] (5) Construct a deep learning network model, use the training set samples as input and the corresponding 3D coordinates of the pacemaker points as labels, and train the entire network model;

[0013] (6) By inputting the test set samples into the trained network model, the location of the ectopic cardiac pacemaker where ventricular premature beats occur can be predicted.

[0014] Furthermore, the specific implementation method of step (2) is as follows: for the surface ECG data of any channel, firstly, wavelet filtering is used to perform denoising processing on it, then an ECG signal sequence of a cardiac cycle and containing a QRS complex is selected therefrom, and finally, the ECG signal sequences of all channels are spliced ​​in the order of the positions of the 12 leads to obtain the time domain information of the ECG signal.

[0015] Furthermore, the specific implementation method of step (3) is as follows: first, the 12-lead surface electrocardiogram data is subjected to one-dimensional Fourier transform lead by lead to obtain the frequency domain expression of the electrocardiogram signal of each channel, which is composed of a complex expression consisting of amplitude and phase, and then the amplitude characteristics and phase characteristics thereof are extracted and spliced ​​in the order of the positions of the 12 leads to obtain the frequency domain information of the electrocardiogram signal.

[0016] Furthermore, the deep learning network model adopts a CNN-BiLSTM-SelfAttention architecture, which includes three independent feature extraction modules, a feature fusion module and an artificial neural network module, wherein the three feature extraction modules are used to extract features of the ECG time domain signal, the amplitude and phase information obtained after Fourier transform, respectively, the feature fusion module is used to fuse the extracted features, and the artificial neural network module is used to convert the fused features into the three-dimensional coordinates of the cardiac ectopic pacemaker.

[0017] Furthermore, the feature extraction module is composed of a convolution layer Conv1, a maximum pooling layer M1, a convolution layer Conv2, a maximum pooling layer M2, a convolution layer Conv3, a maximum pooling layer M3, a convolution layer Conv4, a maximum pooling layer M4, a bidirectional LSTM layer BiLstm1, a bidirectional LSTM layer BiLstm2, a discard layer D1 and a self-attention mechanism layer connected in sequence from input to output.

[0018] Furthermore, the feature fusion module is composed of a Concatenate layer, a Flatten layer and a discard layer D2 connected in sequence from input to output, and the artificial neural network module is composed of five fully connected layers Dense1 to Dense5 connected in sequence.

[0019] Furthermore, the convolution layer Conv1 contains 32 one-dimensional convolution kernels of size 3×12 with a step size of 2; the convolution layer Conv2 contains 64 one-dimensional convolution kernels of size 3×32 with a step size of 2; the convolution layer Conv3 contains 128 one-dimensional convolution kernels of size 3×64 with a step size of 1; the convolution layer Conv4 contains 256 one-dimensional convolution kernels of size 3×128 with a step size of 1; the pooling window lengths of the maximum pooling layers M1 to M4 are both 2, and the step sizes are both 2; the number of hidden neurons in the bidirectional LSTM layer BiLstm1 is 40, and the number of hidden neurons in the bidirectional LSTM layer BiLstm2 is 16, and the output value of each node will be returned; the discard layer D1 randomly selects 10% of the network weight parameters to be reset to 0.

[0020] Furthermore, the Concatenate layer is used to concatenate the three sets of input features to obtain a new feature vector that integrates time domain information and frequency domain information. The Flatten layer is used to perform a one-dimensional operation on the feature vector. The discard layer D2 randomly selects 35% of the network weight parameters and resets them to 0 to prevent the model from overfitting due to too many features. The number of neurons in the fully connected layers Dense1 to Dense5 are 256, 128, 64, 32, and 3, respectively.

[0021] Furthermore, the process of training the network model in step (5) is as follows:

[0022] 5.1 Initialize model parameters, including bias and weight of each layer, learning rate, optimization method and maximum number of iterations;

[0023] 5.2 Input the training set samples into the model, the model forward propagates the output corresponding prediction results, and calculates the loss function between the prediction results and the labels;

[0024] 5.3 Based on the loss function, the model parameters are continuously updated using the gradient descent method until the loss function converges or the maximum number of iterations is reached, and the training is completed.

[0025] Furthermore, the optimization method adopts an adaptive motion estimation algorithm Adam, and the loss function adopts a mean square error.

[0026] The present invention creatively applies the classic model of natural language processing in the field of artificial intelligence to the processing of electrocardiogram time series signals, extracts the frequency domain information of the electrocardiogram signal through Fourier transform, integrates the frequency domain features and time domain features, and uses the fusion of multiple features to improve the judgment accuracy of the model, and finally realizes the function of predicting the three-dimensional coordinates of the abnormal ventricular excitation point through the 12-lead electrocardiogram on the surface of the body. Starting from the data-driven perspective, the present invention realizes the precise position inference of the pacemaker point of the premature ventricular contraction patient, which solves the current deficiency that doctors can only judge the approximate area in clinical practice; on the other hand, compared with the clinical mapping system, the present invention is a non-invasive method, which is safer, more convenient and faster, and has great advantages in clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the flow of the method for locating premature ventricular beats of the present invention.

[0028] Figure 2 This is a schematic diagram of the CNN-BiLSTM-SelfAttention network model structure of the present invention.

[0029] Figure 3 It is a three-dimensional display diagram of the coordinate values ​​of the ectopic pacemaker predicted by the present invention and the actual coordinate values.

[0030] Figure 4 It is a schematic diagram of the error between the coordinate values ​​of the ectopic pacemaker predicted by the present invention and the true coordinate values. DETAILED DESCRIPTION

[0031] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0032] like Figure 1 As shown, the ventricular premature beat localization method of the present invention integrating time domain and frequency domain feature deep learning includes the following steps:

[0033] (1) Obtain 12-lead body surface potential data of the human body.

[0034] In clinical operation, the patient needs to stick on a medical 12-lead electrode patch to collect 12-lead body surface potential data, and perform subsequent data processing operations to predict the three-dimensional coordinates of the cardiac ectopic pacemaker. The training process of the present invention uses a 12-lead electrocardiogram corresponding to a standard human ventricular finite element model, which can obtain relatively complete electrocardiogram data of ventricular regional points, making the prediction of the entire ventricular region more accurate, and then marking each set of electrocardiogram data corresponding to the patient's three-dimensional coordinates of the cardiac ectopic pacemaker as a label.

[0035] (2) Preprocess the collected 12-lead body surface potential data.

[0036] The 12-lead data collected clinically often contains noise information, and the data needs to be filtered. This step can be done by smoothing the signal after wavelet filtering, or on a public software platform such as ECGViewer; then professionals screen the QRS complex of the above signal and roughly select an ECG signal within a cardiac cycle. In this embodiment, the ECG signal sampling frequency is 1000hz, the signal time length is selected to be 512ms, and the size of each lead time series data is 512*1.

[0037] (3) The processed 12-lead electrocardiogram is subjected to Fourier transform to obtain frequency domain features.

[0038] The preprocessed ECG signals are first standardized lead by lead, and then one-dimensional Fourier transform is performed on each lead. The expression of the frequency domain signal after the transformation is composed of a complex expression. The amplitude and phase in the complex expression are split into two features, and the size is also 512*1; the amplitude feature is standardized, and the phase feature is not processed otherwise.

[0039] (4) Splice data and divide data sets.

[0040] The feature data obtained by the above processing are spliced ​​according to the position order of the 12 leads (I II III aVR aVL aVF V1V2 V3 V4 V5 V6). Each sample has three sets of feature vectors: time series signal vector, frequency domain amplitude vector, and frequency domain phase vector, all of which are 6144*1 in size. Finally, there are 3485 sets of samples in total, which are divided into training set, validation set, and test set in a ratio of 8:1:1 and input into the network model for training.

[0041] (5) Build a network model.

[0042] The network model of the present invention is a multi-input single-output neural network, which is mainly composed of three independent feature extraction modules, a layer of feature fusion module and an artificial neural network module; Figure 2As shown, the feature extraction network is mainly composed of three modules: a one-dimensional convolution module, a bidirectional LSTM module and a self-attention mechanism module. In the field of natural language processing, these three types of neural networks have achieved good applications, but their emphases are different. The convolutional neural network is a feedforward neural network whose neurons can respond to surrounding units within a part of the coverage range. It is the most commonly used neural network in the field of deep learning; compared with the other two modules, the convolutional neural network structure is simpler, but it does not consider the connection between each unit before and after, which makes it ignore the global information when extracting features. The LSTM network is a time recursive neural network, which is suitable for processing and predicting important events with relatively long intervals and delays in time series, and can well capture the global features of ECG signals. The introduction of the self-attention mechanism module can enable the entire network model to focus on the time series fragments containing important feature signals from the long ECG time series signal, thereby improving the feature extraction efficiency and accuracy of the entire model. The present invention creatively integrates three modules, overcomes the limitations of each type of nerve, combines their advantages, and improves the positioning accuracy of ectopic pacemakers.

[0043] The one-dimensional convolution module contains four one-dimensional convolution layers Conv1~Conv4. Each convolution layer is connected to a maximum pooling layer. Each convolution layer defines a filter with a length of 3. After four one-dimensional convolutions, the number of convolution kernels increases to 32, 64, 128, and 256 respectively. The pooling window length of the maximum pooling layer is 2, the step size is 2, and the feature signal size of the module is finally outputted. 50*256.

[0044] The LSTM module is composed of two bidirectional LSTM layers BiLstm1~BiLstm2. The bidirectional LSTM is composed of forward LSTM and backward LSTM. It can capture the dependency relationship between the forward and backward time series signals. The LSTM framework is composed of the input word X at time t. t , cell state C t , temporary cell state Hidden state h t 、Forget gate t 、Memory Gate t , output gate o t The calculation process of LSTM can be summarized as follows: by forgetting the information in the cell state and memorizing new information, the information useful for subsequent moment calculations can be transmitted, while the useless information is discarded, and the hidden state h is output at each time step. t , where forgetting, memory and output are determined by the hidden state h at the previous moment t-1 and the current input X t The calculated forget gate f t 、Memory Gate t , output gate o tTo control. After the bidirectional LSTM performs the above operations on the previous input vector and the reversed vector respectively, the latent vectors are concatenated to obtain a new latent vector, which contains the new feature information of the signal. In this embodiment, the number of hidden neurons in BiLstm1 is 40, and the number of hidden neurons in BiLstm2 is 16. The output value of each node will be returned. The output of the LSTM module enters the self-attention mechanism module after passing through a discard layer D1. The discard layer randomly selects 10% of the network weight parameters and resets them to 0.

[0045] The self-attention mechanism mainly solves the long-distance dependency problem by calculating the influence of time series signals on each other. The main calculation process of this module is: the size of the input feature in this step is feature length * feature dimension; first initialize three random matrices and biases b1 and b2, multiply two of them with the input matrix respectively to obtain new matrices query and key, and then expand the two matrices by one dimension respectively, and calculate the matrix h related to the similarity of each part of the input feature according to the formula h = tanh (query + key + b1); then, multiply the matrix h by the third random weight matrix, add the bias b2, and then re-form the matrix e of the size of the original feature length * the original feature length, and normalize the matrix e to obtain the self-attention matrix a; finally, multiply the matrix a by the input feature matrix to obtain a new feature matrix with a data dimension size of 6*32. The above process is expressed by the formula:

[0046] h=tanh(input*w query +input*w key +b1)

[0047] e=input*w e

[0048]

[0049] output = a * input

[0050] The feature fusion module includes a concatenate layer, a flatten layer and a discard layer D2. After performing the above feature extraction operations on the time domain and frequency domain features respectively, three independent feature expressions can be obtained. The concatenate function is used to concatenate the three features in the second dimension to obtain a new feature vector that combines the time domain information and the frequency domain information. A flatten layer is added to perform a one-dimensional operation on the new feature vector, and a discard layer D2 with a discard probability of 0.35 is added to prevent overfitting of the model due to too many features.

[0051] The artificial neural network module consists of five fully connected Dense layers connected in sequence to locate the three-dimensional coordinates of the origin of ventricular premature beats. Through five layers of fully connected convolution operations and nonlinear functions, we can use the fused feature vector to predict the three-dimensional coordinates of the ectopic pacemaker of ventricular premature beats.

[0052] (6) Model training.

[0053] The three sets of feature samples (ECG signal, amplitude, phase) and labels (three-dimensional coordinates of the pacemaker) are input into the network model one by one to obtain high-dimensional features that integrate time domain and frequency domain information. The output high-dimensional features are then regressed and predicted by the fully connected layer to obtain the predicted coordinates of the abnormal ventricular excitation point. Then, the Adam optimizer is selected to optimize the model parameters by minimizing the square difference (or mean square error) between the true coordinate value P (x, y, z) and the predicted coordinate value. The calculation formula is as follows. Finally, the positioning model of the ventricular pacemaker point that is closest to the true coordinate can be obtained.

[0054]

[0055] (7) Input the test set samples into the trained neural network model to test the performance of the network and verify the generalization of the model. The predicted three-dimensional coordinate results will be displayed in the same coordinate system as the actual three-dimensional coordinates, and the prediction accuracy of the model can be intuitively observed.

[0056] (8) Simulation experiment results.

[0057] We mainly conducted two experiments. The first experiment was to predict the test set data to evaluate the model. The second experiment divided the 12-lead ECG data according to the medical clinical ventricular division areas to test the model's prediction accuracy for regional pacemaker points.

[0058] The results of experiment 1 are as follows Figure 3 and Figure 4 As shown, we can intuitively see that the model of the present invention can still well predict the three-dimensional coordinates of the pacemaker for untrained data, and the average prediction error is 3.44 mm, which meets the accuracy standard required for clinical ablation surgery.

[0059] The results of Experiment 2 are shown in Table 1. We divide the ventricle into nine regions: right ventricular outflow tract (RVOT), left ventricular outflow tract (LVOT), tricuspid valve (TV), mitral valve (MV), left ventricle (LV), right ventricle (RV), epicardium (OUT), left ventricular top (LVSUMMIT), and left and right ventricular junction (CRUX). The model of the present invention analyzes and predicts the ECG signals from the nine regions respectively. The right side of Table 1 shows the error between the predicted coordinate value and the true coordinate value. It can be seen that the model has a good prediction effect for each heart region, especially for the left and right ventricular outflow tract where ventricular premature beats often occur. The prediction error can be around 1mm, which is close to the accuracy of the clinical invasive mapping system CARTO (0.7mm) to a large extent, and is far better than the manual judgment of doctors.

[0060] Table 1

[0061]

[0062]

[0063] The above experiments use Tensorflow's Keras platform to build neural networks. The algorithm language is python, which can be used in conjunction with multiple program development environments. The computer CPU used is AMD Ryzen 5 5600X 6-CoreProcessor, the host memory is 32GB RAM, and the graphics card model is NVIDIA GeForce RTX 3070.

[0064] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made to the present invention by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. A ventricular premature beat location method integrating time domain and frequency domain feature deep learning, comprising the following steps: (1) Collect 12-lead surface electrocardiogram data of a large number of patients with different ventricular premature beats, and mark the three-dimensional coordinates of the cardiac ectopic pacemaker corresponding to each set of electrocardiogram data; (2) preprocessing the 12-lead surface electrocardiogram data to obtain time domain information of the electrocardiogram signal; (3) performing a one-dimensional Fourier transform on the 12-lead surface electrocardiogram data lead by lead to obtain frequency domain information of the electrocardiogram signal including amplitude and phase; (4) combining the time domain information and frequency domain information of the ECG signal as a set of samples, and dividing all samples into a training set, a validation set, and a test set; (5) Construct a deep learning network model, use the training set samples as input and the corresponding 3D coordinates of the pacemaker points as labels, and train the entire network model; The deep learning network model adopts a CNN-BiLSTM-SelfAttention architecture, which includes three independent feature extraction modules, a feature fusion module and an artificial neural network module, wherein the three feature extraction modules are used to extract features from the ECG time domain signal, the amplitude and phase information obtained after Fourier transform, the feature fusion module is used to fuse the extracted features, and the artificial neural network module is used to convert the fused features into the three-dimensional coordinates of the cardiac ectopic pacemaker; The feature extraction module is composed of a convolution layer Conv1, a maximum pooling layer M1, a convolution layer Conv2, a maximum pooling layer M2, a convolution layer Conv3, a maximum pooling layer M3, a convolution layer Conv4, a maximum pooling layer M4, a bidirectional LSTM layer BiLstm1, a bidirectional LSTM layer BiLstm2, a discard layer D1 and a self-attention mechanism layer connected in sequence from input to output; the feature fusion module is composed of a Concatenate layer, a Flatten layer and a discard layer D2 connected in sequence from input to output, and the artificial neural network module is composed of 5 fully connected layers Dense1 to Dense5 connected in sequence; (6) By inputting the test set samples into the trained network model, the location of the ectopic cardiac pacemaker where ventricular premature beats occur can be predicted.

2. The method for locating premature ventricular beats according to claim 1, characterized in that: The specific implementation method of step (2) is as follows: for the surface ECG data of any channel, firstly, wavelet filtering is used to perform denoising processing on it, then an ECG signal sequence of a cardiac cycle and containing a QRS complex is selected therefrom, and finally, the ECG signal sequences of all channels are spliced ​​in the order of the positions of the 12 leads to obtain the time domain information of the ECG signal.

3. The method for locating premature ventricular beats according to claim 1, characterized in that: The specific implementation method of step (3) is as follows: first, one-dimensional Fourier transform is performed on the 12-lead surface electrocardiogram data lead by lead to obtain the frequency domain expression of the electrocardiogram signal of each channel, which is composed of a complex expression composed of amplitude and phase, and then the amplitude characteristics and phase characteristics are extracted and spliced ​​in the order of the positions of the 12 leads to obtain the frequency domain information of the electrocardiogram signal.

4. The method for locating premature ventricular beats according to claim 1, characterized in that: The convolution layer Conv1 contains 32 one-dimensional convolution kernels of size 3×12 with a step size of 2; the convolution layer Conv2 contains 64 one-dimensional convolution kernels of size 3×32 with a step size of 2; the convolution layer Conv3 contains 128 one-dimensional convolution kernels of size 3×64 with a step size of 1; the convolution layer Conv4 contains 256 one-dimensional convolution kernels of size 3×128 with a step size of 1; the pooling window lengths of the maximum pooling layers M1 to M4 are both 2, and the step sizes are both 2; the number of hidden neurons in the bidirectional LSTM layer BiLstm1 is 40, and the number of hidden neurons in the bidirectional LSTM layer BiLstm2 is 16, and the output value of each node will be returned; the discard layer D1 randomly selects 10% of the network weight parameters to be reset to 0.

5. The method for locating premature ventricular beats according to claim 1, characterized in that: The Concatenate layer is used to concatenate the three sets of input features to obtain a new feature vector that integrates time domain information and frequency domain information. The Flatten layer is used to perform a one-dimensional operation on the feature vector. The discard layer D2 randomly selects 35% of the network weight parameters and resets them to 0 to prevent the model from overfitting due to too many features. The number of neurons in the fully connected layers Dense1 to Dense5 are 256, 128, 64, 32, and 3, respectively.

6. The method for locating premature ventricular beats according to claim 1, characterized in that: The process of training the network model in step (5) is as follows: 5.1 Initialize model parameters, including bias and weight of each layer, learning rate, optimization method and maximum number of iterations; 5.2 Input the training set samples into the model, the model forward propagates the output corresponding prediction results, and calculates the loss function between the prediction results and the labels; 5.3 Based on the loss function, the model parameters are continuously updated using the gradient descent method until the loss function converges or the maximum number of iterations is reached, and the training is completed.

7. The method for locating premature ventricular beats according to claim 6, characterized in that: The optimization method adopts the adaptive motion estimation algorithm Adam, and the loss function adopts the mean square error.

Images