Arrhythmia intelligent multi-classification device for dynamic electrocardio

Abstract

The invention discloses an arrhythmia intelligent multi-classification device for dynamic electrocardio. The arrhythmia intelligent multi-classification device comprises a data acquisition preprocessing module, an arrhythmia intelligent multi-classification neural network construction and training module, an automatic coding module and an integration module. The data acquisition and preprocessing module is used for performing feature extraction and preprocessing on electrocardiogram data, an electrocardiosignal data set is constructed and comprises features of the electrocardiosignal data and category label values corresponding to the features of the electrocardiosignal data, and the arrhythmia intelligent multi-classification neural network comprises a plurality of convolution blocks, a normalization layer and an activation function. And training a neural network by using the electrocardiosignal data set to obtain an intelligent arrhythmia multi-classification model. And the automatic coding module takes the characteristic value of the electrocardiosignal data as input to obtain a corresponding classification result. The integration module compares the output of the two models and outputs a final classification result, the classification accuracy can be improved for arrhythmia categories with small data volume, and the device can provide diagnosis reference and basis for clinicians.

Description

Technical Field

[0001] The present invention belongs to the field of medical signal processing, and in particular relates to an intelligent multi-classification device for arrhythmias in dynamic electrocardiograms. Background Art

[0002] ECG signal preprocessing, feature extraction and diagnostic classification are three important components of arrhythmia classification. The accuracy of feature extraction affects the final output of medical equipment. Normal ECG signals contain many waveforms, which are divided by the zero potential of excitation, mainly including P wave, QRS complex, T wave and other parts. All waves correspond to specific cardiac activities and electrophysiological periods. For the test of QRS waveform, the main methods include filter method, template matching method, wavelet transform method, neural network method, etc. Awni Y. Hannun et al. published a paper on January 7, 2019 (https: / / doi.org / 10.1038 / s41591-018-0268-3) to conduct a comprehensive evaluation of end-to-end deep learning methods for ECG analysis. The method was verified based on an independent test data set annotated by a consensus committee of board-certified practicing cardiologists, which confirmed that the method has high diagnostic performance similar to that of cardiologists, indicating that convolutional neural networks based on deep learning have broad application prospects in the field of ECG processing. There are many algorithms nowadays. Yao et al. proposed a model for arrhythmia classification, which integrated CNN and LSTM layers based on VGG-Net. The model achieved an overall classification F1 of 77.3% on a 9-classification ECG signal test set. He et al. proposed a model composed of a residual convolutional network and a bidirectional LSTM layer. The model achieved an overall F1 of 80.6% in 9 arrhythmia classifications. Yao et al. proposed to introduce an attention module after the CNN layer. This model is an effective method for detecting paroxysmal arrhythmias. Wang et al. proposed to use GAN and residual network for ECG arrhythmia detection, which solved the problem of data set imbalance and improved the classification accuracy. Hou et al. proposed a network based on LSTM autoencoder and used SVM to classify these signals. The overall accuracy of the model was high, but the classification accuracy was low for arrhythmia categories with less data. Summary of the invention

[0003] In view of the problems existing in the prior art, the present invention provides an intelligent multi-classification device for arrhythmia of dynamic electrocardiogram, which can classify the dynamic electrocardiogram signals detected and acquired into four types: normal, atrial fibrillation, premature beats and other abnormalities, and can output the time period information of the abnormal electrocardiogram signals, so as to facilitate the location and further processing of the abnormal signals. The more important advantage is that the present invention uses the established deep neural network module and the automatic encoder module to process the data in multiple modules, and uses the voting method of the integrated module to compare the classification results of the two modules to obtain the final classification result, so as to improve the accuracy of diagnosis.

[0004] To solve the above technical problems, the present invention provides the following technical solutions: an intelligent multi-classification device for arrhythmia of dynamic electrocardiogram, comprising a data acquisition preprocessing module, an intelligent multi-classification neural network construction and training module for arrhythmia, an automatic encoding module, and an integration module;

[0005] The data acquisition preprocessing module preprocesses the collected ECG data to extract feature values ​​and construct an ECG signal data set including the features of the ECG signal data and its corresponding category label values;

[0006] The module for constructing and training an intelligent multi-classification neural network for arrhythmia takes the characteristic value of the electrical signal data as input and the corresponding label value as output, constructs and trains an intelligent multi-classification neural network for arrhythmia, and obtains an intelligent multi-classification model for arrhythmia; the intelligent multi-classification neural network for arrhythmia includes a first convolution block and a second convolution block connected in sequence, several convolution blocks with the same structure, a normalization layer, and an activation function layer; when training the neural network, the first convolution block outputs the preliminary features of the electrocardiogram signal, the second convolution block outputs the features of the electrocardiogram signal in an updated state, several convolution blocks with the same structure perform cyclic extraction on the features of the electrocardiogram signal and output the final features, the normalization layer is used to improve the stability of the training and accelerate the convergence, and the activation function layer outputs the category with the highest probability according to the final features as the output of the model;

[0007] The automatic encoding module includes a symmetrical encoder and decoder, which takes the characteristic values ​​of the ECG signal data as input to obtain the corresponding classification results of normal, atrial fibrillation, premature beats or other abnormalities;

[0008] The integration module compares the output results of the arrhythmia intelligent multi-classification neural network construction and training module and the automatic encoding module, and outputs the final arrhythmia classification result.

[0009] Furthermore, the aforementioned intelligent multi-classification device for arrhythmia of dynamic electrocardiogram preprocesses the collected electrocardiogram data including noise removal, filtering and signal enhancement.

[0010] Furthermore, the aforementioned intelligent multi-classification device for arrhythmia for dynamic electrocardiogram, extracting features from collected electrocardiogram data includes extracting time domain features, frequency domain features, and time-frequency domain features.

[0011] Furthermore, the aforementioned first convolution block includes a sequentially connected input layer, a first convolution unit, a first normalization layer, and a first activation function ReLU layer; the input layer serves as the input end of the first convolution block to receive the characteristic values ​​of the ECG signal data output by the data acquisition preprocessing module, and the activation function ReLU layer serves as the output end of the first convolution block to output preliminary features of the ECG signal.

[0012] Furthermore, the aforementioned second convolution block includes a second convolution unit, a second normalization layer, a second activation function ReLU layer, a first Dropout unit, a first LSTM unit, and a third convolution unit connected in sequence; the second convolution unit serves as the input end of the second convolution block to receive preliminary features of the ECG signal output by the first convolution block, and the third convolution unit serves as the output end of the second convolution block to output updated ECG signal features.

[0013] Furthermore, in the aforementioned several convolution blocks with the same structure, each convolution block includes a normalization layer A, an activation function ReLU layer A, a convolution unit A, a normalization layer B, an activation function ReLU layer B, a Dropout unit, an LSTM unit, and a convolution unit B which are sequentially connected; the normalization layer A of the first convolution block in the several convolution blocks with the same structure receives the ECG signal features of the updated state output by the third convolution unit of the second convolution block, and the convolution unit B of the last convolution block in the several convolution blocks with the same structure serves as the output end to output the final features.

[0014] Furthermore, the expression of the aforementioned normalization layer is as follows:

[0015] y i =BN γ，β (x i )

[0016] In the formula, xi represents the input feature value, γ and β represent the learnable scaling parameter and translation parameter respectively, and BNγ,β(xi) represents the output feature obtained by batch normalization of the input feature.

[0017] Furthermore, the expression of the activation function ReLU layer mentioned above is as follows:

[0018]

[0019] In the formula, x represents the eigenvalue.

[0020] Furthermore, the aforementioned automatic encoding module is configured to perform the following actions: the encoder encodes the characteristic values ​​of the ECG signal data, and then inputs the encoded characteristic values ​​into a hidden layer composed of several neurons for mapping; the decoder decodes the characteristic values ​​output by the hidden layer and outputs the classification results of the automatic encoding module.

[0021] Furthermore, the aforementioned integrated module uses a voting method to compare the classification results output by the arrhythmia intelligent multi-classification neural network construction and training module and the automatic encoding module, and when the result is abnormal, outputs the time information of the abnormal signal.

[0022] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0023] The present invention realizes the automatic analysis of dynamic ECG signals. Based on the wearable dynamic ECG data of patients with arrhythmia, after preprocessing by the data loading module, the constructed deep neural network and automatic encoder are used to classify the ECG data, which are specifically divided into four categories: normal, atrial fibrillation, premature beats and other abnormalities. Subsequently, the classification results of the two models are compared by the voting mechanism through the integration module to draw the final conclusion. When the result is displayed as abnormal, the system can synchronously provide the time location information of the abnormal signal, and has strong noise resistance and interpretability. The device provides diagnostic reference and basis for clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural diagram of intelligent multi-classification neural network for arrhythmia.

[0025] Figure 2 This is the structure diagram of the automatic encoding module

[0026] Figure 3 This is the working principle diagram of the automatic encoding module. DETAILED DESCRIPTION

[0027] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0028] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.

[0029] The present invention provides an intelligent multi-classification device for arrhythmia of dynamic electrocardiogram, comprising a data acquisition preprocessing module, an intelligent multi-classification neural network construction and training module for arrhythmia, an automatic encoding module, and an integration module;

[0030] The data acquisition preprocessing module preprocesses the collected ECG data to extract feature values ​​and construct an ECG signal data set including the features of the ECG signal data and its corresponding category label values;

[0031] An intelligent multi-classification neural network construction and training module for arrhythmia takes the characteristic values ​​of the electrical signal data as input and the corresponding label values ​​as output, constructs and trains an intelligent multi-classification neural network for arrhythmia to obtain an intelligent multi-classification model for arrhythmia; an automatic encoding module, including a symmetrical encoder and decoder, takes the characteristic values ​​of the ECG signal data as input, and obtains the corresponding classification results of normal, atrial fibrillation, premature beats or other abnormalities;

[0032] The integration module compares the output results of the arrhythmia intelligent multi-classification neural network construction and training module and the automatic encoding module, and outputs the final arrhythmia classification result.

[0033] The data group of the data acquisition preprocessing module is divided into a training set and a validation set. The data group is sorted by sample size, and data is generated in batches during the training of the deep neural network module and the automatic encoding module and added to the training set to realize the recycling of data. This embodiment is about automatically analyzing the electrocardiogram data of patients with arrhythmia. First, after preprocessing by the data processing module, the data becomes cleaner and more usable. Then, two main modules are used: a deep neural network module and an automatic encoder module to classify the electrocardiogram data into four categories: normal, atrial fibrillation, premature beats or other abnormalities. Then, the classification results of the two modules are compared by the voting method in the integration module, and the final classification diagnosis result is obtained. This process is intended to provide doctors with more accurate and efficient diagnostic support.

[0034] The data processing module is responsible for preprocessing long-term ECG signals and dividing the data into training sets and validation sets for subsequent model training and evaluation. In one example, ECG data from Cardiogram, a manufacturer of ECG monitoring instruments, is evaluated. Their equipment can collect ECG signal data from patients through smart watches or smartphone applications. The dataset used contains 50,000 ECG records from 25,000 patients, each of which lasts for about 7 to 10 days. The age distribution of these patients is wide, ranging from 18 to 80 years old, of which 60% are male and 40% are female. ECG signal data is obtained from the database provided by Cardiogram and converted into a format suitable for algorithm processing. ECG signal data is preprocessed, including operations such as noise removal, filtering, and signal enhancement to ensure the quality and reliability of the data. Each ECG record is labeled according to the patient's medical history and clinical diagnosis results, such as normal, arrhythmia, myocardial ischemia, etc. The labeled dataset is divided into training sets and validation sets, usually 80% of the data is used for training the model and 20% is used for validating the model. Features are extracted from each ECG record, including time domain features (such as heart rate variability, QRS wave width), frequency domain features (such as power spectral density) and time-frequency domain features (such as wavelet transform coefficients).

[0035] The extracted feature values ​​and label values ​​are sorted in a certain order and packaged into an array. Then, the data processing efficiency is improved through batch processing, and new batches of data are gradually generated during training to achieve data recycling.

[0036] The data processed by the data processing module can provide high-quality data samples for the training of deep learning models and provide basic support for the accuracy and reliability of arrhythmia classification and diagnosis algorithms.

[0037] refer to Figure 1 The arrhythmia intelligent multi-classification neural network includes a first convolution block and a second convolution block connected in sequence, several convolution blocks with the same structure, a normalization layer and an activation function layer; when training the neural network, the first convolution block outputs the preliminary features of the ECG signal, the second convolution block outputs the features of the ECG signal in an updated state, several convolution blocks with the same structure perform cyclic extraction on the ECG signal features and output the final features, the normalization layer (Batch Normalization) is used to improve the stability of training and accelerate convergence, and the activation function layer outputs the category with the highest probability according to the final features as the output of the model;

[0038] Among them, the first convolution block includes a sequentially connected input layer, a first convolution unit, a first normalization layer, and a first activation function ReLU layer; the input layer serves as the input end of the first convolution block to receive the characteristic values ​​of the ECG signal data output by the data acquisition preprocessing module, and the activation function ReLU layer serves as the output end of the first convolution block to output preliminary features of the ECG signal.

[0039] The second convolution block includes a second convolution unit, a second normalization layer, a second activation function ReLU layer, a first Dropout unit, a first LSTM unit, and a third convolution unit connected in sequence; the second convolution unit serves as the input end of the second convolution block to receive the preliminary features of the ECG signal output by the first convolution block, and the third convolution unit serves as the output end of the second convolution block to output the ECG signal features of the updated state.

[0040] Among several convolution blocks with the same structure, each convolution block includes a normalization layer A, an activation function ReLU layer A, a convolution unit A, a normalization layer B, an activation function ReLU layer B, a Dropout unit, an LSTM unit, and a convolution unit B which are sequentially connected; the normalization layer A of the first convolution block among several convolution blocks with the same structure receives the ECG signal features of the updated state output by the third convolution unit of the second convolution block, and the convolution unit B of the last convolution block among several convolution blocks with the same structure serves as an output end to output the final features.

[0041] The above activation functions include ReLU activation function and softmax activation function. ReLU activation function introduces nonlinearity and activates features. The normalization layer (Batch Normalization) standardizes the convolution results and speeds up the training process. The Dropout layer can randomly disable some neurons to reduce the risk of overfitting. The LSTM layer is used to process sequence data and retain and update state information. These layers operate in a loop to help the model understand the data more comprehensively and improve the accuracy of classification. Each convolution block performs the same operation so that the network can better learn the characteristics of the data, thereby improving the performance of the overall model.

[0042] Specifically, the batch normalization layer and the activation function layer play an important role in the deep neural network module. The normalization layer is used to improve the stability of model training and accelerate the convergence speed. It standardizes the features of each small batch of data so that the mean and variance of the input features are close to the standard normal distribution, thereby accelerating the training process of the model. The batch normalization layer calculates the activation mean and standard deviation of each feature for each small batch of data, and then standardizes each feature. The standardized feature values ​​can center the activation values ​​around zero and scale them to a standard deviation of one, which helps to accelerate the training process of the model and improve the performance of the model. The mathematical model of the batch normalization operation is expressed as follows:

[0043] y i =BN γ，β (x i )

[0044] Among them, xi represents the input feature value, γ and β represent the learnable scaling parameter and translation parameter respectively, and BNγ,β(xi) represents the output feature obtained by batch normalization operation on the input feature.

[0045] The activation function layer introduces nonlinear characteristics, allowing the neural network model to learn more complex function mapping relationships. It performs nonlinear transformations on the output of neurons, thereby increasing the expressiveness and adaptability of the model. The activation function layer introduces nonlinear properties by performing nonlinear transformations on the output of neurons, allowing the neural network model to learn more complex function mappings, not just linear transformations. Activation functions include ReLU function and Softmax function. The mathematical expression of the ReLU function is:

[0046]

[0047] The ReLU function keeps positive inputs unchanged, which helps avoid the vanishing gradient problem, and introduces sparsity by mapping negative inputs to zero, which helps with feature selection and sparse representation. The Softmax function converts the original output into a probability distribution and selects the category with the highest probability as the output.

[0048] The max pooling layer reduces the spatial dimension of the data by sliding a fixed-size window over the input data and selecting the maximum value in the window as the output value. The max pooling layer introduces spatiotemporal invariance, which improves the model's robustness to the input data and generalization performance by retaining the most significant features in each window and ignoring minor information.

[0049] The regularization layer is used to reduce the overfitting of the model and prevent the model from overfitting on the training set, thereby improving the generalization ability of the model on the test set. Common regularization techniques include Dropout, L1 regularization, and L2 regularization. L1 regularization penalizes the absolute value of the model's parameters, prompting the model to learn simpler features and reducing the risk of overfitting. L2 regularization penalizes the square of the model's parameters, reducing the weight of features with larger parameter values ​​and reducing the degree of overfitting of the model. The introduction of the regularization layer can effectively improve the generalization ability of the deep learning model, making it more robust and reliable on unseen data.

[0050] refer to Figure 2 ,The autoencoder module is used to learn useful representations of input data and improve the model's ability to understand and classify ECG signal features. This module includes encoder and decoder parts, which map the input data to a low-dimensional representation through a series of neural network layers and decompress it to the original input data. The autoencoder module has high noise robustness and can handle noisy or incomplete ECG signals.

[0051] The integration module integrates the classification results of the deep neural network module and the autoencoder module, and uses the voting method to select the final classification diagnosis result. Through the integration module, the algorithm can fully utilize the advantages of the deep neural network module and the autoencoder module to obtain higher classification accuracy, robustness and generalization performance.

[0052] Output layer: The classified diagnosis results and the time information of abnormal signals are output in the final output layer.

[0053] The automatic encoding module includes an encoder and a decoder. The encoder and the decoder are symmetrical and both contain multiple hidden layers composed of neurons.

[0054] Combination Figure 3As shown in Figure 1, the encoder part maps the input data to a low-dimensional representation, and the decoder part maps the encoded representation back to the original input space. The decoder is symmetrical to the encoder and contains multiple hidden layers composed of neurons.

[0055] The auto-encoding module makes up for the lack of interpretability of the deep neural network module to a certain extent. Visualization tools or feature response analysis tools can be used to view the importance of different ECG signal features to classification decisions, which is critical for providing diagnostic evidence. In addition, the auto-encoder model has high noise robustness and can handle noisy or incomplete ECG signals.

[0056] The deep neural network model and the autoencoder model are trained separately, and then the voting method is used in the integrated module to select the category with the highest number of votes as the final classification result. Compared with the autoencoder model, the deep neural network module can capture abstract features in the data, performs well when processing data with highly nonlinear relationships, and is good at processing large training data sets; compared with the deep neural network module, the autoencoder module has advantages such as good noise processing capabilities. Therefore, the final integrated module has higher accuracy, robustness, and generalization performance, while retaining a certain degree of interpretability.

[0057] The diagnosis result is normal, atrial fibrillation, premature beats or other abnormalities. When the diagnosis result is abnormal, the output of the integrated module also includes time information of the abnormal signal.

[0058] To further demonstrate the effectiveness and significance of the arrhythmia multi-classification model of the present invention for ECG signals, according to the provided validation set, the deep neural network module of the present invention is used to give two data, validation loss and validation accuracy, wherein the validation loss is the loss index of the module on the validation data set, which is used to evaluate the generalization performance of the model. The lower the value, the better the performance of the model on the validation set; the validation accuracy is the accuracy index of the module on the validation data set. The higher the value, the higher the classification accuracy of the module on the validation set. 10% of the input data set is selected as the validation set, and the deep neural network module is used to classify the ECG signals. The classification effect is shown in Table 1, a comparison table of the ECG signal results of the deep neural network module training set and the validation set. Through horizontal comparison, it is found that the generalization performance of the validation set in the present invention is better and the accuracy is higher.

[0059] Table 1

[0060]

[0061]

[0062] The present invention improves and adds appropriate error handling mechanisms to handle possible error situations. For example, a try-except block is used to capture and handle exceptions. A more advanced ResNet model, such as ResNet50 or ResNet101, is used to improve network performance.

[0063] At the same time, the parameters of the convolutional layer, such as the kernel size, the number of kernels, and the stride, are adjusted to adapt to different datasets and tasks.

[0064] Add regularization techniques, such as Batch Normalization or dropout, to improve the generalization ability of the model and prevent overfitting.

[0065] The invention supports two datasets, The 2017PhysioNet / CinC Challenge and MIT-BH, to meet different experimental needs. A simple and clear command line interface allows model training with just a few lines of code. It uses deep learning technology to improve ECG classification accuracy, which has significant advantages over traditional methods. Based on Python and Keras, it ensures seamless integration with other scientific computing libraries. Adjust the parameters of the output layer, such as the number of output categories and activation functions, to adapt to different classification tasks. This improves the accuracy.

[0066] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. An intelligent multi-classification device for arrhythmia in dynamic electrocardiogram, characterized in that: It includes a data acquisition preprocessing module, an intelligent multi-classification neural network construction and training module for arrhythmia, an automatic encoding module, and an integration module; The data acquisition preprocessing module preprocesses the collected ECG data to extract feature values ​​and construct an ECG signal data set including the features of the ECG signal data and its corresponding category label values; The module for constructing and training an intelligent multi-classification neural network for arrhythmia takes the characteristic value of the electrical signal data as input and the corresponding label value as output, constructs and trains an intelligent multi-classification neural network for arrhythmia, and obtains an intelligent multi-classification model for arrhythmia; the intelligent multi-classification neural network for arrhythmia includes a first convolution block and a second convolution block connected in sequence, several convolution blocks with the same structure, a normalization layer, and an activation function layer; when training the neural network, the first convolution block outputs the preliminary features of the electrocardiogram signal, the second convolution block outputs the features of the electrocardiogram signal in an updated state, several convolution blocks with the same structure perform cyclic extraction on the features of the electrocardiogram signal and output the final features, the normalization layer is used to improve the stability of the training and accelerate the convergence, and the activation function layer outputs the category with the highest probability according to the final features as the output of the model; The automatic encoding module includes a symmetrical encoder and decoder, which takes the characteristic values ​​of the ECG signal data as input to obtain the corresponding classification results of normal, atrial fibrillation, premature beats or other abnormalities; The integration module compares the output results of the arrhythmia intelligent multi-classification neural network construction and training module and the automatic encoding module, and outputs the final arrhythmia classification result.

2. The intelligent multi-classification device for arrhythmia according to claim 1, characterized in that: The preprocessing of collected ECG data includes noise removal, filtering and signal enhancement.

3. The intelligent multi-classification device for arrhythmia according to claim 1, characterized in that: Feature extraction of collected electrocardiogram data includes extracting time domain features, frequency domain features, and time-frequency domain features.

4. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The first convolution block includes a sequentially connected input layer, a first convolution unit, a first normalization layer, and a first activation function ReLU layer; the input layer serves as the input end of the first convolution block to receive the characteristic values ​​of the ECG signal data output by the data acquisition preprocessing module, and the activation function ReLU layer serves as the output end of the first convolution block to output preliminary features of the ECG signal.

5. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The second convolution block includes a second convolution unit, a second normalization layer, a second activation function ReLU layer, a first Dropout unit, a first LSTM unit, and a third convolution unit connected in sequence; the second convolution unit serves as the input end of the second convolution block to receive the preliminary features of the ECG signal output by the first convolution block, and the third convolution unit serves as the output end of the second convolution block to output the ECG signal features of the updated state.

6. The intelligent multi-classification device for arrhythmia according to claim 5, characterized in that: Among several convolution blocks with the same structure, each convolution block includes a normalization layer A, an activation function ReLU layer A, a convolution unit A, a normalization layer B, an activation function ReLU layer B, a Dropout unit, an LSTM unit, and a convolution unit B which are sequentially connected; the normalization layer A of the first convolution block among several convolution blocks with the same structure receives the ECG signal features of the updated state output by the third convolution unit of the second convolution block, and the convolution unit B of the last convolution block among several convolution blocks with the same structure serves as an output end to output the final features.

7. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The expression of the normalization layer is as follows: yes i =BN γ,β (x) i ) In the formula, xi represents the input feature value, γ and β represent the learnable scaling parameter and translation parameter respectively, and BNγ,β(xi) represents the output feature obtained by batch normalization of the input feature.

8. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The expression of the activation function ReLU layer is as follows: Where x represents the eigenvalue.

9. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The automatic encoding module is configured to perform the following actions: the encoder encodes the characteristic values ​​of the ECG signal data, and then inputs the encoded characteristic values ​​into the hidden layer for mapping; the decoder decodes the characteristic values ​​output by the hidden layer composed of several neurons, and outputs the classification results of the automatic encoding module.

10. The intelligent multi-classification device for arrhythmia in dynamic electrocardiogram according to claim 1, characterized in that: The integrated module uses voting method to compare the classification results output by the arrhythmia intelligent multi-classification neural network construction and training module and the automatic encoding module. When the result is abnormal, the time information of the abnormal signal is output.

Images