Classification Method and Device for Electrocardiogram Signals

Abstract

The present invention provides a method and apparatus for classifying electrocardiogram (ECG) signals. The method includes: inputting an ECG signal waveform into a classification network model to obtain a classification result, where the classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after performing band-pass filtering on the ECG signal waveform. The waveform feature extraction module is used to extract statistical features of the ECG signal waveform. The deep features and the classification statistical features are concatenated and then input into the fully connected layer; generating a classification report according to the statistical features and the classification result; and outputting the classification report. This method uses the deep features and statistical features of the ECG signal waveform for classification, improving the accuracy of the classification result. At the same time, by adding statistical features, the extracted feature signals are more interpretable, thereby improving the interpretability of the classification result of the classification network model.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine, and particularly to a method for classifying electrocardiogram signals. Background Art

[0002] In recent years, the incidence and fatality rate of cardiovascular diseases have been continuously increasing. Before the onset of cardiovascular disease patients, arrhythmia often occurs. Electrocardiogram signals can directly reflect the heart rate of patients, which is beneficial to the diagnosis and treatment of various heart diseases. Therefore, the detection and classification of electrocardiogram signals have important clinical significance for the diagnosis and treatment of cardiovascular diseases.

[0003] Currently, the main electrocardiogram signal classification methods are classification methods based on deep learning. This method can rely on the powerful feature extraction ability of neural networks and a larger model capacity to classify electrocardiogram signals.

[0004] Currently, for the classification method based on deep learning, its classification result accuracy is low. Summary of the Invention

[0005] An embodiment of the present invention provides a method for classifying electrocardiogram signals to solve the problem of low accuracy of classification results in the classification method based on deep learning.

[0006] In a first aspect, an embodiment of the present invention provides a method for classifying electrocardiogram signals, including:

[0007] Inputting the electrocardiogram signal waveform into a classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after band-pass filtering the electrocardiogram signal waveform. The waveform feature extraction module is used to extract statistical features of the electrocardiogram signal waveform. The deep features and the statistical features are concatenated and then input into the fully connected layer;

[0008] Generating a classification report according to the statistical features and the classification result;

[0009] Outputting the classification report.

[0010] In a possible implementation, the deep feature extraction network includes multiple convolutional layers;

[0011] The first convolutional layer of the deep feature extraction network is used to perform band-pass filtering on the electrocardiogram signal waveform;

[0012] The remaining convolutional layers of the deep feature extraction network are used to extract deep features from the filtered electrocardiogram signal waveform.

[0013] In a possible implementation, the first convolutional layer of the depth feature extraction network includes a Sinc convolutional layer, a pooling layer, and a normalization layer;

[0014] The remaining convolutional layers of the depth feature extraction network all include: a convolutional layer, a pooling layer, and a normalization layer.

[0015] In a possible implementation, the waveform feature extraction module is used to perform feature analysis on the electrocardiogram signal waveform to obtain a plurality of initial statistical features, and the SHAP-Value algorithm is used to screen the plurality of initial statistical features to obtain at least one of the statistical features.

[0016] In a possible implementation, the statistical features include one or more of the following features: morphological features, rhythm features.

[0017] In a possible implementation, generating a classification report according to the statistical features and the classification result includes:

[0018] Analyzing and processing the statistical features according to the normal value range and abnormal value range of the statistical features to obtain a diagnostic index;

[0019] Generating the classification report according to the diagnostic index and the classification result.

[0020] In a possible implementation, the method further includes:

[0021] Obtaining a training data set, where the training data set includes a plurality of training samples, and the training samples include electrocardiogram signal waveforms;

[0022] Using the training data set to train a preset first classification network model to obtain the classification network model, and the parameters in the classification network model are updated by the gradient descent method. The first classification network model includes a depth feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module.

[0023] In a possible implementation, the classification module uses softmax classification.

[0024] In a second aspect, an embodiment of the present invention provides a classification device for electrocardiogram signals, including:

[0025] An input module for inputting an electrocardiogram (ECG) signal waveform into a classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after band-pass filtering the ECG signal waveform. The waveform feature extraction module is used to extract statistical features of the ECG signal waveform. The deep features and the statistical features are concatenated and then input into the fully connected layer;

[0026] A generation module for generating a classification report according to the statistical features and the classification result;

[0027] An output module for outputting the classification report.

[0028] In a possible implementation, the deep feature extraction network includes multiple convolutional layers;

[0029] The first convolutional layer of the deep feature extraction network is used to perform band-pass filtering on the ECG signal waveform;

[0030] The remaining convolutional layers of the deep feature extraction network are used to extract deep features from the filtered ECG signal waveform.

[0031] In a possible implementation, the first convolutional layer of the deep feature extraction network includes a Sinc convolutional layer, a pooling layer, and a normalization layer;

[0032] All the remaining convolutional layers of the deep feature extraction network include: a convolutional layer, a pooling layer, and a normalization layer.

[0033] In a possible implementation, the waveform feature extraction module is used to perform feature analysis on the ECG signal waveform to obtain multiple initial statistical features, and uses the SHAP-Value algorithm to screen the multiple initial statistical features to obtain at least one of the statistical features.

[0034] In a possible implementation, the statistical features include one or more of the following features: morphological features, rhythm features.

[0035] In a possible implementation, the generation module is specifically used for:

[0036] Analyzing and processing the statistical features according to the normal value range and abnormal value range of the statistical features to obtain a diagnostic index;

[0037] Generating the classification report according to the diagnostic index and the classification result.

[0038] In a possible implementation, it further includes an acquisition module, and the acquisition module is used for:

[0039] Obtain a training data set, where the training data set includes a plurality of training samples, and the training samples include electrocardiogram signal waveforms;

[0040] Use the training data set to train a preset first classification network model to obtain the classification network model. The parameters in the classification network model are updated using the gradient descent method. The first classification network model includes a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module.

[0041] In a possible implementation manner, the classification module uses softmax classification.

[0042] In a third aspect, an embodiment of the present invention provides a classification device for electrocardiogram signals, including:

[0043] At least one processor and a memory;

[0044] The memory stores computer execution instructions;

[0045] The at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the classification method for electrocardiogram signals provided in the first aspect.

[0046] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, in which computer execution instructions are stored, and when the computer execution instructions are executed by a processor, they are used to implement the classification method for electrocardiogram signals provided in the first aspect.

[0047] A classification method and device for electrocardiogram signals provided by the present invention use a classification network model to extract deep features and statistical features from the input electrocardiogram signal waveform, and classify according to the deep features and statistical features, improving the accuracy of the classification result. At the same time, when the classification network model provided by the present invention extracts deep features from electrocardiogram signal data, it first performs band-pass filtering on the electrocardiogram signal waveform. When the existing network model extracts deep features, it will not filter the electrocardiogram signal. Compared with the existing method, the deep features extracted by the method of the present invention have better interpretability. In addition, the present invention adds statistical features. Compared with the classification method that directly uses deep features, the extracted feature signals are more interpretable, thereby improving the interpretability of the classification result of the classification network model. Description of the Drawings

[0048] The drawings here are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0049] Figure 1Schematic diagram of the classification network model provided by the embodiments of the present invention;

[0050] Figure 2 Schematic diagram of the deep feature extraction network provided by the embodiments of the present invention;

[0051] Figure 3 Flowchart of a method for classifying electrocardiogram signals provided by Embodiment 1 of the present invention;

[0052] Figure 4 Flowchart of a method for training the classification network model provided by Embodiment 2 of the present invention;

[0053] Figure 5 Flowchart of generating an electrocardiogram signal report provided by Embodiment 3 of the present invention;

[0054] Figure 6 Schematic diagram of the structure of an electrocardiogram signal classification device provided by Embodiment 4 of the present invention;

[0055] Figure 7 A classification device for electrocardiogram signals provided by Embodiment 5 of the present invention.

[0056] Through the above-mentioned drawings, specific embodiments of the present invention have been shown, and more detailed descriptions will be given later. These drawings and textual descriptions are not intended to limit the scope of the concept of the present invention in any way, but to illustrate the concept of the present invention to those skilled in the art by referring to specific embodiments. Detailed Description of the Embodiments

[0057] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0058] The detection and classification of electrocardiogram signals have important clinical significance for the diagnosis and treatment of cardiovascular diseases. For the existing method of classification based on artificial features, that is, relying on prior knowledge in the medical field to extract time-domain and frequency-domain features of electrocardiogram signals for classification, the classification results are highly interpretable, but there are many interference factors due to the use of artificial features during classification; for the classification method based on deep learning, its network model lacks interpretability, and the black box of the neural network is difficult to be trusted by users such as doctors and patients.

[0059] To solve the problems of the prior art, an embodiment of the present invention provides a method for classifying electrocardiogram signals, which uses a trained classification network model for classification. This classification network model is different from the existing classification network models.

[0060] Figure 1 It is a schematic structural diagram of the classification network model provided by the embodiment of the present invention. As Figure 1 shown, the classification network model includes a deep feature extraction network 101, a waveform feature extraction module 102, a fully connected layer 103, and a classification module 104.

[0061] The deep feature extraction network 101 is used to extract deep features after band-pass filtering the electrocardiogram signal waveform. The deep feature extraction network 101 includes a convolutional module with a band-pass filter. The band-pass filter learns the upper and lower cut-off frequencies and extracts the electrocardiogram frequency band that contributes to classification diagnosis from the original electrocardiogram signal. For different cases, the influence of the electrocardiogram frequency band in the electrocardiogram signal is different. Exemplarily, for the diagnosis of coronary heart disease and heart failure diseases, the influence of different electrocardiogram frequency bands in the electrocardiogram signal on the diagnosis of specific causes is different, so it is necessary to select the part with greater influence for analysis.

[0062] The waveform feature extraction module 102 is used to extract the statistical features of the electrocardiogram signal waveform. These statistical features usually have clear physical meanings, and the results obtained by classifying through these statistical features are more interpretable. Mainly relying on prior knowledge in the medical field, feature extraction in the time domain and frequency domain of the electrocardiogram signal is performed for classification. These statistical features include one or more of the following features: morphological features, rhythm features.

[0063] The fully connected layer 103 fuses the deep knowledge features and statistical features, and inputs the fused features into the classification module 104.

[0064] The classification module 104 classifies the features input by the fully connected layer 103 to obtain the classification result of the electrocardiogram signal. Specifically, this classification module can use softmax classification. The softmax classifier is a linear classifier that can calculate the probabilities for all classification labels.

[0065] For the classification network model provided by the present invention, when electrocardiogram signal data is input into this network model, by extracting the deep features and statistical features of the electrocardiogram signal waveform, and using the deep features and statistical features of the electrocardiogram signal waveform for classification, the features used by this classification network model are multi-angle and comprehensive, thereby being able to improve the accuracy of the classification result. In addition, due to the addition of statistical features, compared with the classification method directly using deep features, the extracted feature signals are more interpretable, thereby improving the interpretability of the classification result of this classification network model.

[0066] The following combines Figure 2 to elaborate in detail on the deep feature extraction network in the classification network model of the embodiments of the present invention.

[0067] Figure 2 The structure diagram of the deep feature extraction network provided for the embodiments of the present invention is as follows. As Figure 2 shown, the deep feature extraction network includes multiple convolutional layers. Among them, the first layer of the deep extraction network is used to perform band-pass filtering on the electrocardiogram signal waveform. Exemplarily, Sinc convolution can be used to implement band-pass filtering. Specifically, the Sinc convolution includes a Sinc convolutional layer, a pooling layer, and a normalization layer. The remaining convolutional layers of the deep feature extraction network are used to extract deep features from the filtered electrocardiogram signal waveform. The remaining convolutional layers of the deep feature extraction network are all ordinary convolutions, specifically including a convolutional layer, a pooling layer, and a normalization layer.

[0068] The input electrocardiogram signal data first undergoes Sinc convolution in the first layer of the deep feature extraction network. The Sinc convolution is a Sinc band-pass filtering convolution. The band-pass filter in this Sinc band-pass filtering convolution has a clear physical meaning, and its specific expression is as follows:

[0069] y[n] = x[n] * g[n, θ]

[0070] where g is the band-pass filter and θ is the learnable parameter, which is the upper and lower cut-off frequencies in the band-pass filter. Expanding the band-pass filter in the frequency domain form, we can obtain:

[0071] G[f, f1, f2] = rect(f / 2f2) - rect(f / 2f1)

[0072] where f1 and f2 are the upper and lower cut-off frequencies respectively.

[0073] When calculating the deep features extracted from the electrocardiogram signal data using the deep feature extraction network, the calculation is based on the time domain. Therefore, the band-pass filter in the deep feature extraction network is a time-domain filter. Specifically, it can be obtained by Fourier transform of a frequency-domain filter, and its expression is as follows:

[0074] g[n, f1, f2] = 2f2sinc(2πf2n) - 2f1sinc(2πf1n)

[0075] When extracting deep features from the input electrocardiogram signal data, first, the first layer Sinc convolutional layer of the deep feature extraction network performs band-pass filtering on the electrocardiogram signal waveform, and then the remaining convolutional layers of the deep feature extraction network extract deep features from the filtered electrocardiogram signal waveform to obtain the deep features as Figure 2 shown.

[0076] Existing deep feature extraction networks usually consist of multiple ordinary convolutional layers as described above, that is, they do not filter the electrocardiogram signal waveform first. In this embodiment, the first Sinc convolutional layer of the deep feature extraction network performs band-pass filtering on the electrocardiogram signal waveform, so that the obtained deep features can have better interpretability.

[0077] According to Figure 1 and Figure 2 the network model shown, the method for classifying electrocardiogram signals using this model will be described in detail below in combination with Figure 3 ,

[0078] Figure 3 FIG. Figure 3 is a flowchart of a method for classifying electrocardiogram signals provided in Embodiment 1 of the present invention. As

[0079] shown, the method for classifying the electrocardiogram signal includes the following steps.

[0080] Input the electrocardiogram signal waveform into the classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. Figure 2 Input the electrocardiogram signal waveform into the classification network model. The deep feature extraction network extracts the deep features of the electrocardiogram signal waveform by the method provided in the embodiment as

[0081] shown. The waveform feature extraction module extracts the statistical features of the electrocardiogram signal waveform according to the electrocardiogram signal waveform and the existing feature analysis methods in the electrocardiogram field. The extracted deep features and statistical features are concatenated and then input into the fully connected layer for fusion, and the fused features are input into the classification module to classify the features. When classifying, a softmax classifier can be used to classify the input fused features. Specifically, the probability values of different classes are calculated for each feature value through the softmax classifier. Among them, the sum of the probabilities of all feature possibilities is 1, and the classification result is obtained according to the probability value of each feature. Exemplarily, assuming that the obtained feature values are 3.5, 8, and 7.4 respectively, then the probability values of three class possibilities can be calculated through the softmax classifier as: 0.18, 0.42, and 0.39.

[0082] The statistical features extracted by the waveform feature extraction module include one or more of the following features: morphological features, rhythm features. The electrocardiogram (ECG) signal is not a single waveform and is usually composed of the superposition of P wave, Q wave, U wave, QRS complex, ST segment, T wave, etc. Morphological features are used to describe the waveform features of various waves included in the ECG signal. The waveform features may include the amplitude, duration, and spectrum of the wave, etc., and may also include the P-R interval, P-P interval, average value, variance, maximum value, minimum value of the signal, etc. Rhythm features are used to describe the regular features of cardiac activities, and the rhythm features include one or more of sinus rhythm, atrial escape rhythm, atrioventricular node escape rhythm, and ventricular escape rhythm, etc.

[0083] In one implementation, the waveform feature extraction module can perform feature analysis on the ECG signal waveform to obtain multiple initial statistical features, and use these initial statistical features as the final statistical features.

[0084] In another implementation, the waveform feature extraction module can perform feature analysis on the ECG signal waveform to obtain multiple initial statistical features, and use the interpretable machine learning algorithm (SHapley Additive explanation value, abbreviated as SHAP-Value) to screen the multiple initial statistical features to obtain at least one statistical feature. For any prediction sample, the SHAP-Value is the value obtained by the proportion of the weight of each feature in the sample. Through this value, the contribution degree of the features in each sample can be reflected, and at the same time, the positive and negative correlations of the influence can also be reflected.

[0085] In this embodiment, the influence of the same statistical feature on different etiologies is different. Therefore, for different etiologies, it is necessary to use SHAP-Value to screen the multiple initial statistical features, that is, to evaluate the influence and importance of each feature on the classification result of the etiology, and obtain the features that are effective for the classification result of the etiology.

[0086] Exemplarily, for the electrocardiogram signal waveform of a coronary heart disease patient, assuming that the multiple initial statistical features extracted by the waveform feature extraction module include P wave, Q wave, ST segment, T wave, P-R interval, P-P period, and QT interval, the SHAP-Value corresponding to each feature is obtained according to the value of each initial statistical feature, and the contribution degree of the feature corresponding to the value is determined according to the value. Specifically, the larger the value, the greater the corresponding feature contribution degree. By setting an initial threshold, the initial statistical features are screened according to the threshold and the SHAP-Value corresponding to each initial feature, and the features effective for the cause classification result are obtained. According to the existing knowledge in the electrocardiogram field, we can know that the electrocardiogram signal mainly shows ST segment, T wave, Q wave, and QT interval under this cause. Therefore, the effective features obtained through the above screening can be one or more of ST segment, T wave, Q wave, and QT interval.

[0087] S302. Generate a classification report according to the statistical features and the classification result.

[0088] In a possible implementation manner, the classification result obtained by processing through the classification network model can provide a diagnostic reference for medical patients. Therefore, the statistical features and the classification result are generated into a classification report in text form.

[0089] In another possible implementation manner, according to the statistical features and the relevant knowledge in the electrocardiogram field, the normal value range and the interpretation of the abnormal value range of the statistical feature are obtained. Specifically, the normal value range and the abnormal value range of each statistical feature are set according to empirical values, and the value of the statistical feature extracted during the classification process is compared with the normal value range and the abnormal value range. According to the comparison result, a diagnostic index is determined. The diagnostic index can reflect some bases for obtaining the classification result. The diagnostic index is usually an index that doctors and patients can directly understand. A classification report is generated according to the diagnostic index and the classification result.

[0090] S303. Output the classification report.

[0091] The classification device can directly display the classification report through a display screen, or send the classification report to other devices for display or printing. The classification device can also save the classification report locally for subsequent query.

[0092] In this embodiment, an electrocardiogram (ECG) signal is input into a classification network model. The deep feature network and waveform extraction module in the classification network model are used to extract features from the ECG signal, obtaining deep features and statistical features, and finally generating a classification report through processing of the above features. On the one hand, when the classification network model adopted by this method performs deep feature extraction, it uses Sinc band-pass filtering convolution to perform band-pass filtering on the ECG signal waveform, enabling the obtained deep features to have better interpretability and also improving the accuracy of classification. On the other hand, this method extracts the deep features and statistical features of the ECG signal waveform and uses them for classification, improving the accuracy of the classification result.

[0093] The following will combine Figure 4 to elaborate in detail on the training of the classification network model. Figure 4 It is a flowchart of the training method for the classification network model provided in the second embodiment of the present invention. As Figure 4 shown, the training of the classification network model specifically includes the following steps.

[0094] S401: Obtain a training data set, which includes multiple training samples.

[0095] The training data set is the MIT-BIH Arrhythmia Database, which includes more than 4,000 24-hour periodic dynamic ECG signal data of 47 test individual units, totaling 109,500 heartbeats, of which abnormal heartbeats account for about 30%. The database has unified naming for different arrhythmia types, specifically including: normal heartbeat, left bundle branch block, right bundle branch block, atrial premature beat, ventricular premature beat, fusion wave, pacing heartbeat, ventricular escape beat, atrial escape beat, ventricular fibrillation, junctional premature beat, supraventricular premature beat, and unclassified heartbeat. Specifically, each sample in this training set also includes a classification result, and its classification result includes one or more of the above different arrhythmia types in this database.

[0096] Obtain training samples from this data set and input the samples into a preset first classification network model for training.

[0097] S402: Use the training data set to train the preset first classification network model to obtain a classification network model.

[0098] The first classification network model includes a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. Among them, the functions of each part are the same as those in the above classification network model and will not be elaborated here.

[0099] The samples obtained from the dataset are first passed through a deep feature network model and a waveform feature module to extract deep features and statistical features. The extraction methods are the same as those in the above embodiments and will not be elaborated here.

[0100] The extracted deep features and statistical features are concatenated. Subsequently, using the concatenated features, the deep feature extraction network, fully connected layer, and classification module in the first preset network model are trained. During the training process, the parameters in the model are updated using the gradient descent method. Specifically, gradient descent is a type of iterative method. Briefly speaking, it is a method for finding the minimum of an objective function. That is, when solving the minimum value of the loss function, the gradient descent method can be used to iteratively solve step by step to obtain the minimized loss function and model parameter values. By using the gradient descent method to iteratively train the network model until the loss value of the classification result obtained by the classification module reaches the minimum compared with the classification result of the input electrocardiogram signal data, the training is ended to obtain the classification network model.

[0101] In this embodiment, the training samples obtained from the MIT - BIH dataset are input into the preset first classification network model for training. The model parameters are updated through iterative training using the gradient descent method. After meeting the iterative conditions, the training is stopped to obtain the classification network model. This method improves the accuracy of the classification result of electrocardiogram signal data through the training of the model.

[0102] Based on the embodiment, the embodiment of the present invention provides a method for generating an electrocardiogram signal report, which is used to elaborate on step S303 in Embodiment 1. Figure 5 This is a flowchart of generating an electrocardiogram signal report provided in Embodiment 3 of the present invention. As Figure 5 shown, the generation of this electrocardiogram signal report specifically includes the following steps.

[0103] S501, Analyze and process the statistical features according to the normal value range and abnormal value range of the statistical features to obtain diagnostic indicators.

[0104] Set the normal value range and abnormal value range of each statistical feature according to empirical values. Compare the values of the statistical features extracted during the classification process with this normal value range and abnormal value range. Determine the diagnostic indicators based on the comparison results. The diagnostic indicators can reflect some bases for this classification result. The diagnostic indicators are usually indicators that doctors and patients can directly understand, thus making the classification result more interpretable and readable.

[0105] Exemplarily, for the electrocardiogram signal data of patients with coronary heart disease, when the ST segment depression is 0.1 - 0.4 mV, it is diagnosed as ischemic ST segment change.

[0106] S502. Generate a classification report based on the diagnostic indicators and classification results.

[0107] This classification report outputs the diagnostic indicators and classification results in text form.

[0108] In this embodiment, the corresponding diagnostic indicators are obtained by analyzing and processing the statistical features obtained by the network classification model. A classification report is generated based on the diagnostic indicators and classification results, and finally this classification report is output in text form. The report obtained by this method is convenient for doctors and patients to understand.

[0109] Figure 6 The structural schematic diagram of an electrocardiogram signal classification device provided in Embodiment 4 of the present invention. As Figure 6 shown, a classification device 60 for electrocardiogram signals includes: an input module 601, a generation module 602, and an output module 603.

[0110] The input module 601 is configured to input the electrocardiogram signal waveform into the classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after band-pass filtering the electrocardiogram signal waveform. The waveform feature extraction module is used to extract the statistical features of the electrocardiogram signal waveform. The deep features and statistical features are concatenated and then input into the fully connected layer;

[0111] The generation module 602 is configured to generate a classification report according to the statistical features and classification results;

[0112] The output module 603 is configured to output the classification report.

[0113] In a possible implementation manner, the deep feature extraction network includes multiple convolutional layers;

[0114] The first convolutional layer of the deep feature extraction network is used to perform band-pass filtering on the electrocardiogram signal waveform;

[0115] The remaining convolutional layers of the deep feature extraction network are used to extract deep features from the filtered electrocardiogram signal waveform.

[0116] In a possible implementation manner, the first convolutional layer of the deep feature extraction network includes a Sinc convolutional layer, a pooling layer, and a normalization layer;

[0117] The remaining convolutional layers of the deep feature extraction network all include: a convolutional layer, a pooling layer, and a normalization layer.

[0118] In a possible implementation manner, the waveform feature extraction module is used to perform feature analysis on the electrocardiogram signal waveform to obtain multiple initial statistical features, and the SHAP-Value algorithm is used to screen the multiple initial statistical features to obtain at least one statistical feature.

[0119] In a possible implementation, the statistical features include one or more of the following features: morphological features, rhythm features.

[0120] In a possible implementation, the generation module is specifically configured to:

[0121] Analyze and process the statistical features according to the normal value range and abnormal value range of the statistical features to obtain a diagnostic index;

[0122] Generate a classification report according to the diagnostic index and the classification result.

[0123] In a possible implementation, it further includes an acquisition module, and the acquisition module is used to:

[0124] Acquire a training data set, where the training data set includes multiple training samples, and the training samples include electrocardiogram signal waveforms;

[0125] Use the training data set to train a preset first classification network model to obtain a classification network model, and the parameters in the classification network model are updated by the gradient descent method. The first classification network model includes a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module.

[0126] The device provided in this embodiment can be used to execute the method steps of the above method embodiment. The specific implementation manners and technical effects are similar and will not be elaborated here.

[0127] Figure 7 This embodiment of the present invention provides a classification device for electrocardiogram signals, including:

[0128] At least one processor 701 and a memory 702;

[0129] The memory 702 stores computer execution instructions;

[0130] At least one processor 701 executes the computer execution instructions stored in the memory 702, so that at least one processor 701 executes the above electrocardiogram signal classification method.

[0131] The specific implementation process of the processor 701 can refer to the above method embodiment. The specific implementation manners and technical effects are similar and will not be elaborated here.

[0132] This embodiment of the present invention provides a computer-readable storage medium, in which computer execution instructions are stored. When the computer execution instructions are executed by a processor, they are used to implement the steps of the above method embodiment. The specific implementation manners and technical effects are similar and will not be elaborated here.

[0133] Other embodiments of the invention will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The invention is intended to cover any variations, uses, or adaptations of the invention following the general principles of the invention and including known or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are pointed out by the following claims.

[0134] It should be understood that the invention is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for classifying electrocardiogram signals, characterized in that, Including: Input the electrocardiogram (ECG) signal waveform into a classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after band-pass filtering the ECG signal waveform. The waveform feature extraction module is used to perform feature analysis on the ECG signal waveform to obtain multiple initial statistical features, and use the SHAP-Value algorithm to screen the multiple initial statistical features to obtain at least one of the statistical features. The deep features and the statistical features are concatenated and then input into the fully connected layer; Generate a classification report according to the statistical features and the classification result; Output the classification report; The deep feature extraction network includes multiple convolutional layers; the first convolutional layer of the deep feature extraction network is used to perform band-pass filtering on the ECG signal waveform; the remaining convolutional layers of the deep feature extraction network are used to extract deep features from the filtered ECG signal waveform; the first convolutional layer of the deep feature extraction network includes a Sinc convolutional layer, a pooling layer, and a normalization layer; the remaining convolutional layers of the deep feature extraction network each include: a convolutional layer, a pooling layer, and a normalization layer.

2. The method according to claim 1, wherein The statistical features include one or more of the following features: morphological features, rhythm features.

3. The method according to claim 1, wherein The generating a classification report according to the statistical features and the classification result includes: Analyze and process the statistical features according to the normal value range and abnormal value range of the statistical features to obtain a diagnostic index; Generate the classification report according to the diagnostic index and the classification result.

4. The method according to any one of claims 1 to 3, characterized in that The method further includes: Obtain a training data set, where the training data set includes multiple training samples, and the training samples include ECG signal waveforms; Use the training data set to train a preset first classification network model to obtain the classification network model, and update the parameters in the classification network model using the gradient descent method. The first classification network model includes a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module.

5. An electrocardiogram signal classification device, characterized in that, Including: An input module, which is used to input the ECG signal waveform into a classification network model to obtain a classification result. The classification network model includes: a deep feature extraction network, a waveform feature extraction module, a fully connected layer, and a classification module. The deep feature extraction network is used to extract deep features after band-pass filtering the ECG signal waveform. The waveform feature extraction module is used to perform feature analysis on the ECG signal waveform to obtain multiple initial statistical features, and use the SHAP-Value algorithm to screen the multiple initial statistical features to obtain at least one of the statistical features. The deep features and the statistical features are concatenated and then input into the fully connected layer; A generating module, which is used to generate a classification report according to the statistical features and the classification result; An output module, which is used to output the classification report; The deep feature extraction network includes multiple convolutional layers; the first convolutional layer of the deep feature extraction network is used to perform band-pass filtering on the electrocardiogram signal waveform; the remaining convolutional layers of the deep feature extraction network are used to perform deep feature extraction on the filtered electrocardiogram signal waveform; the first convolutional layer of the deep feature extraction network includes a Sinc convolutional layer, a pooling layer, and a normalization layer; the remaining convolutional layers of the deep feature extraction network each include: a convolutional layer, a pooling layer, and a normalization layer.

6. A classification device for electrocardiogram signals, characterized in that, Comprising: At least one processor and a memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the classification method of the electrocardiogram signal according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Computer-executable instructions are stored in the computer-readable storage medium, and when the computer-executable instructions are executed by a processor, the classification method of the electrocardiogram signal according to any one of claims 1-4 is implemented.

Images