An ECG signal classification system and method based on deep learning neural networks
By employing a multi-source unsupervised domain adaptation method using multi-branch deep learning neural networks, the challenge of cross-individual classification in the ECG dataset was solved, achieving better generalization performance and classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2021-12-29
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies struggle to effectively utilize multiple ECG datasets for cross-individual classification, and traditional methods exhibit poor generalization performance, making them difficult to apply effectively across different datasets.
A multi-branch deep learning neural network is adopted, including a common feature extraction module, a domain-invariant branch module, and a domain-specific branch module. By minimizing the classification loss and domain adaptation loss, and combining prior knowledge for feature matching and prediction fusion, multi-source unsupervised domain adaptation is achieved.
This improves the generalization performance of the ECG signal classification model, enabling the effective application of learned knowledge on unknown target datasets, thereby enhancing classification accuracy and robustness.
Smart Images
Figure CN116432070B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ECG signal classification technology using deep learning algorithms, and more specifically, to an ECG signal classification system and method based on deep learning neural networks. Background Technology
[0002] According to the World Health Organization (WHO), cardiovascular disease is the leading cause of death worldwide, accounting for approximately one-third of all deaths annually. Cardiac arrhythmias are the most common cardiovascular disease. As a non-invasive diagnostic tool, electrocardiography (ECG) is one of the most effective tools for doctors to identify arrhythmias; in fact, most arrhythmias can be diagnosed solely based on ECG. However, analyzing each heartbeat is necessary to detect cardiac abnormalities, but this process is cumbersome and time-consuming. Furthermore, interpreting ECG recordings is challenging and typically only achievable by cardiologists. Therefore, there is a need to design an effective automated ECG classification tool. However, automated classification of ECG signals is a challenging problem. The morphological and temporal characteristics of ECG signals vary significantly among different patients under different physiological conditions. Therefore, for the same disease, different patients may exhibit different patterns. Moreover, due to the difficulty of sample labeling, the number of samples available for classification model learning is very limited.
[0003] ECG describes the electrical fluctuations caused by cardiac depolarization and repolarization. A normal heartbeat cycle consists of five phases: P, Q, R, S, and T. The P wave represents the potential changes during atrial depolarization, the Q, R, and S wave groups represent the potential changes during ventricular depolarization, and the T wave represents rapid ventricular repolarization. If there are abnormalities in the heart, they are usually reflected in the waveform of the heartbeat cycle. Observing changes in the ECG waveform and diagnosing corresponding cardiac abnormalities is the foundation of ECG diagnosis of heart disease. For example, a significant characteristic of ECGs from patients with atrial fibrillation (AF) is the absence of P waves. ECG data comes in various formats, including single-lead, two-lead, six-lead, and twelve-lead ECGs. Twelve-lead ECG data can carry more information, thus helping doctors make a more accurate diagnosis.
[0004] Before the rise of deep learning, the methods for automatically diagnosing ECG were mainly traditional machine learning methods that relied on manually designed features to classify ECG signals. These traditional methods were not very effective.
[0005] In recent years, with the outstanding performance of deep learning in areas such as image classification and speech recognition, more and more researchers have begun to explore the use of deep learning algorithms for ECG data classification. Considering the great success of convolutional neural networks (CNNs) in image classification, many works have applied CNNs to ECG data classification. Given the time-series nature of ECG signals, many works focus on intra-subject ECG classification, where the training and test sets are from the same individuals. Due to the strong individual variability in ECG signals, these methods have poor generalization ability and are difficult to implement practically. To improve generalization performance, a portion of the labeled data from the test set can be used for training, and the remaining samples from the test set can be used to test the model—a so-called patient-specific approach. However, this setup is inconvenient because a portion of the data must be labeled for each new individual. The most general approach would be to train on data from some individuals and test on data from others—a cross-subject approach. However, this setup is very difficult because ECG datasets are limited in sample size and cannot cover all ECG diagnoses. Although there are now more and more publicly available ECG datasets, it is difficult to make effective use of multiple datasets due to the differences in data distribution between different datasets.
[0006] In the field of deep learning, domain adaptation methods have been proposed to address the problem of data distribution discrepancies. Specifically, these methods align the feature distributions of the source domain with those of the target domain using metric-based or adversarial approaches, thereby more effectively utilizing the knowledge learned from the source domain. Summary of the Invention
[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide an ECG signal classification system and method based on deep learning neural networks.
[0008] According to one aspect of the present invention, an ECG signal classification system based on a deep learning neural network is provided, comprising:
[0009] A common feature extraction module, which uses a first convolutional neural network to extract shallow features from the input ECG signal;
[0010] The domain-invariant branch module employs a second convolutional neural network to extract features from the shallow features. By minimizing the classification loss, it performs feature matching for all source domains to obtain a label prediction.
[0011] The domain-specific branch module employs a third convolutional neural network and uses LMMD loss to perform feature matching on the input shallow features between each pair of source and target domains, thereby obtaining n label predictions.
[0012] The hybrid module, based on prior knowledge, assigns different confidence levels to each sub-branch according to the number of samples of various labels that can be obtained by each sub-branch of the domain-invariant branch module and the domain-specific branch module, and combines the n+1 label predictions obtained by the domain-invariant branch module and the domain-specific branch module to obtain the final label prediction.
[0013] Preferably, the first convolutional neural network of the common feature extraction module includes:
[0014] A convolutional layer, comprising a normalization layer and a non-linear layer, wherein the kernel size is 1*7;
[0015] The number of max pooling layers is two.
[0016] A residual module, which is located after the max pooling layer;
[0017] The input to the first convolutional neural network is all ECG signals, and the output is the shallow features of all ECG signal data.
[0018] Preferably, the second convolutional neural network of the domain-invariant branch module includes a domain-invariant feature extractor and a classifier;
[0019] The domain-invariant feature extractor includes a residual module, two max pooling layers, three residual modules with channel attention, and a global average pooling layer.
[0020] The classifier includes a fully connected layer and a Sigmoid layer;
[0021] The input to the second convolutional neural network is the shallow features of the ECG signal, and the output is a prediction of the classification result of the ECG signal data.
[0022] Preferably, the channel attention includes:
[0023] An average pooling layer pools the features of each channel into a single element;
[0024] Two fully connected layers, each followed by a non-linear layer, and a final sigmoid layer, map the pooled features to the weights of the corresponding channels.
[0025] Preferably, the third convolutional neural network of the domain-specific branch module includes: n sub-networks, the structure of which is the same as that of the second convolutional deep neural network, where n is the number of the source domain datasets;
[0026] The input to each sub-network is a pair of shared shallow features from the source and target domain datasets, which are matched using a domain adaptation loss; the output is a prediction of the classification result of the ECG signal data.
[0027] Preferably, the hybrid module combines the n+1 label predictions output by the second and third convolutional networks based on prior statistical information of various label samples in the source domain dataset. in, {zi} represents the predicted values of the n branches of the domain-specific branch; z0 represents the predicted values of the domain-invariant branch; α∈[0,1] is the fusion factor.
[0028] According to a second aspect of the present invention, a method for ECG signal classification based on a deep learning neural network is provided, comprising:
[0029] Obtain the source domain dataset and target domain dataset of the ECG signal;
[0030] Based on the source domain dataset and the target domain dataset, train a common feature extraction network, a domain-invariant branch network, and a domain-specific branch network;
[0031] The ECG signals of the target domain dataset are input into the trained common feature extraction network, domain-invariant branch network, and domain-specific branch network to obtain n+1 predicted classifications.
[0032] The n+1 predicted classifications are fused to obtain the ECG signal classification result.
[0033] Preferably, the step of training a common feature extraction network, a domain-invariant branch network, and a domain-specific branch network based on the source domain dataset and the target domain dataset includes:
[0034] The ECG signals in the source domain dataset are input into a common feature extraction network to extract shallow features of the ECG signals;
[0035] The shallow features are respectively input into the domain-invariant branch network and the domain-specific branch network to obtain one classification and n classifications;
[0036] Calculate the two classification losses for the one category and the n categories;
[0037] The ECG signals in the target domain dataset are input to the domain-specific branch to extract target domain features.
[0038] Calculate the domain adaptation loss between the target domain features and the source domain features;
[0039] Based on the domain adaptation loss and classification loss, the network parameters of the common feature extraction network, the domain invariant branch network, and the domain-specific branch network are updated using the backpropagation method.
[0040] Preferably, fusing the n+1 predicted classifications to obtain the ECG signal classification result includes:
[0041] The number of samples in each source domain dataset is counted according to the category, and then divided by the total number of samples in the corresponding category to obtain the weight of the corresponding classifier for each category sample.
[0042] Based on the weights, the n predictions of the domain-specific branch are merged into a single prediction. The predicted value z0 from the domain-invariant branch is fused with a hyperparameter α: The final predicted value was obtained.
[0043] Preferably, the classification loss function is Focal loss.
[0044] Among them, z i Let l(x) be the predicted value of the model. i ) represents the true label of the sample, and γ is a hyperparameter that controls the penalty strength of the loss function;
[0045] The domain adaptation loss is the LMMD loss.
[0046]
[0047] Where C represents the total number of categories, c represents the category number, and D i Let D represent the i-th source domain. y Denotes the target domain, x i It comes from D i The sample y is from D y The sample, a ic x represents i The weight of the c-th label, b c φ represents the weight of y on the c-th label, and φ represents the kernel function. This represents the regenerated Hilbert space.
[0048] Compared with the prior art, the present invention has the following beneficial effects: The present invention effectively learns the general feature representation of ECG signals through a multi-branch network composed of several branches in the domain-invariant branch and the domain-specific branch and multi-source unsupervised domain adaptation, so that the knowledge learned from the source dataset can be used on the unknown target dataset. Furthermore, through a prior-based classifier fusion strategy, multiple predicted values are organically combined to obtain the final result, which further improves the generalization performance of the model. Attached Figure Description
[0049] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0050] Figure 1 A schematic diagram of the training phase of an ECG signal classification method based on a deep learning neural network, provided as an embodiment of the present invention;
[0051] Figure 2 A schematic diagram of the testing phase of an ECG signal classification method based on a deep learning neural network, as provided in an embodiment of the present invention;
[0052] Figure 3 This invention provides an example of a priori-based weight allocation strategy for an ECG signal classification method based on a deep learning neural network, as an embodiment of the present invention. Detailed Implementation
[0053] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.
[0054] This invention provides an embodiment of an ECG signal classification system based on a deep learning neural network, comprising: a common feature extraction module, a domain-invariant branch module, a domain-specific branch module, and a hybrid module. The common feature extraction module employs a first convolutional neural network to extract shallow features from the input ECG signal; the domain-invariant branch module employs a second convolutional neural network to extract features from the shallow features, implicitly performing feature matching for all source domains by minimizing classification loss to obtain a label prediction; the domain-specific branch module employs a third convolutional neural network, using LMMD loss, to perform feature matching between each pair of source and target domains for the input shallow features, obtaining n label predictions; the hybrid module, based on prior knowledge, assigns different confidence levels to each sub-branch according to the number of samples of each type of label that can be obtained from each sub-branch of the domain-invariant branch module and the domain-specific branch module, and combines the n+1 label predictions obtained from the domain-invariant branch module and the domain-specific branch module to obtain the final label prediction.
[0055] In this embodiment, prior knowledge refers to the assumption that the model's classification performance for each category of samples is proportional to the number of samples available. This embodiment implements multi-source unsupervised domain adaptation. "Multi-source" means that the data used to train the model comes from multiple datasets; "unsupervised" means that there are no labeled samples in the target domain; and "domain adaptation" refers to using methods to reduce the distribution differences across different datasets, so that the knowledge learned from the source datasets can be applied to the target domain. The entire network framework in this embodiment works together to solve the problem of multi-source unsupervised domain adaptation. Implicitly, this means not using any domain adaptation methods. Correspondingly, LMMD loss is explicitly used as a feature matching method in the domain-specific branch. When relying solely on classification loss, to minimize the global classification loss, the model will also select features that are common across all domains as much as possible. The implicit feature matching mentioned here is more about "feature selection," while feature matching in the domain-specific branch is more about "reducing feature distances."
[0056] To better extract common features from ECG signals, this invention provides an embodiment. The common feature extraction module takes the preprocessed ECG signal as input and aims to learn shallow feature representations shared across all datasets. Features learned in shallower layers of a convolutional neural network can generally be considered shallow, common features; features learned in deeper layers carry more task-related information. Based on this, the network structure of the common feature extractor is not complex, containing only three convolutional layers and two residual modules. Each convolutional layer contains a normalization layer and a nonlinear layer; the kernel size is 1*7; the output is the feature representation shared across all datasets.
[0057] To achieve a better classification prediction, this invention provides a preferred embodiment. In this embodiment, the domain-invariant branch takes the features output by the common feature extraction module as input, aiming to learn a feature representation shared across all datasets. For multi-source domain unsupervised domain adaptation problems, aligning domain features is challenging because there are distributional differences not only between the source and target domains but also between the source domains themselves. Due to the specific nature of the ECG multi-source domain unsupervised domain adaptation problem, methods that only align each pair of source and target domains on the domain-specific branch perform poorly because each dataset exhibits data imbalance. Although features may be aligned, the model's classification ability is limited due to the small number of available samples. The domain-invariant branch consists of two parts: a domain-invariant feature extractor and a subsequent classifier. During training, the domain-invariant branch accepts shallow representations of all source domain data. Methods used for single-source domain adaptation offer limited improvement in multi-source domain adaptation. The domain-invariant branch does not use any domain adaptation loss but relies solely on minimizing the classification error of all source domain data to implicitly learn domain-invariant representations.
[0058] To better obtain n classification predictions, this invention provides a preferred embodiment. In this embodiment, the domain-specific branch takes the features output by the common feature extraction module as input, and the goal is to learn the feature representation shared between each pair of source domain datasets and target domain datasets;
[0059] The domain-specific branch contains n sub-networks, representing the number of source domain datasets. Each sub-network consists of two parts: a domain-specific feature extractor and a subsequent classifier. The domain-specific feature extractor contains a residual module, two max-pooling layers, three residual modules with channel attention, and a global average pooling layer. The classifier contains only one fully connected layer and one sigmoid layer. The input to the domain-specific branch is the shallow shared features output from the common feature extractor. Specifically, the sub-network numbered i only accepts samples from the i-th source and target datasets as input. For the source and target domain features output by the domain-specific feature extractor, LMMD loss is used as the feature fitting method. Finally, the output of the fully connected layer is a prediction of the ECG signal classification result.
[0060] The channel attention mechanism includes: (1) an average pooling layer that pools the features of each channel into a single element; (2) two fully connected layers, each followed by a non-linear layer; and (3) a final sigmoid layer that maps the pooled features to the weights of the corresponding channels. The channel attention mechanism makes the model pay more attention to those key features, thereby improving the performance of the model.
[0061] To fuse n+1 predictions into a final prediction, this invention provides a preferred embodiment. The hybrid module synthesizes the n+1 predictions from both the domain-invariant branch and the domain-specific branch to provide the final prediction. First, combine the predicted values {zi} of the n branches of the domain-specific branch: Then the predicted value z0 of the domain-invariant branch and Combined: Where the fusion factor α∈[0,1].
[0062] Furthermore, a prior-based weight allocation method is provided. The performance of the classifiers in the proposed model is inconsistent because each classifier has access to different training data. For a classifier in a domain-specific branch, it can only be trained using data from a specific source and target domain. For a classifier in a domain-invariant branch, it can be trained using all the data. Based on this, a prior-based fusion strategy is proposed. Consider a binary classification problem, setting the number of negative samples to a large constant, and then gradually increasing the number of positive samples from zero. It is assumed that the classification performance of the model is linearly positively correlated with the number of positive samples, provided that positive samples are always less than negative samples. This assumption is not general and counterexamples can be easily found, but it is at least reasonable in ECG classification problems. First, the number of samples in each class within each source domain is calculated, resulting in n vectors V_1, V_2, V_3, ..., V_n. Then the weights are given by the following formula:
[0063]
[0064] Where, ω ic v represents the weight of the i-th classifier for the c-th class. ic This represents the number of class c samples in the i-th dataset.
[0065] Based on the same concept as the above embodiments, the present invention provides an embodiment. An ECG signal classification method based on a deep learning neural network, comprising:
[0066] S100, acquire the source domain dataset and target domain dataset of the ECG signal;
[0067] S200 trains a common feature extraction network, a domain-invariant branch network, and a domain-specific branch network based on the source domain dataset and the target domain dataset.
[0068] S300 inputs the ECG signals of the target domain dataset into the trained common feature extraction network, domain-invariant branch network, and domain-specific branch network to obtain n+1 predicted classifications;
[0069] S400 fuses the n+1 predicted classifications to obtain the ECG signal classification result.
[0070] Based on the above embodiments, the present invention provides a preferred embodiment for performing an ECG signal classification method.
[0071] First, let's introduce the basic settings of this embodiment. There are n labeled source domain datasets X = {x1, ..., xn}. n The objective of this embodiment is to train a model that performs as well as possible on the target domain y, i.e., to achieve more accurate classification of the target domain dataset y.
[0072] This embodiment describes the model training phase. The model training phase generally includes several iterations; this embodiment will focus on one iteration in detail.
[0073] like Figure 1 The diagram shown illustrates a single iteration of the model in this embodiment, including:
[0074] S0: Randomly select a source domain i from the source domain dataset X, and randomly select a batch of sample data {x} from that source domain. i ,l i}
[0075] S1: Randomly select a batch of sample data y from the target domain.
[0076] S2: x i The data is fed into the common feature extractor F to obtain the common shallow feature representation F(x). i ).
[0077] S3: Feed y into the common feature extractor F to obtain the common shallow feature representation F(y).
[0078] S4: F(x) i The feature extractor S0 in the domain-invariant branch is fed into the feature extractor S0 to obtain S0(F(x)). i )).
[0079] S5: Set S0(F(x) i The predicted value z0 is obtained by feeding the data into the classifier C0 in the domain-invariant branch.
[0080] S6: Using x i The tag l i The classification loss is calculated based on the predicted value z0 using Focal loss.
[0081] S7: F(x) i The feature extractor S in the domain-specific branch is fed into. i , obtain S i (F(xi )).
[0082] S8: S i (F(x i The classifier C is fed into the domain-specific branch. i The predicted value z is obtained. i .
[0083] S9: Using x i The tag l i Compared with the predicted value z i The classification loss is calculated based on Focal loss.
[0084] S10: Feed F(y) into the feature extractor S for the domain-specific branch. i , obtain S i (F(y)).
[0085] S11: In order to align the features of the source domain i with those of the target domain, S i (F(x i )) and S i The domain adaptation loss (LMMD loss) is calculated between (F(y)).
[0086] S12: Using the obtained domain adaptation loss and classification loss, update the network parameters based on the backpropagation algorithm.
[0087] like Figure 2 The diagram illustrates the testing phase process of this embodiment. As can be seen, during the testing phase, samples from the target domain are fed into all network branches. The domain-invariant branch will provide a single predicted value z0, while the domain-specific branch will provide n predicted values z0, ..., zn. n The mixing module is responsible for combining the n+1 predicted values to obtain the final prediction.
[0088] Furthermore, a prior-based strategy is employed. Specifically, the classifier weights are allocated based on the proportion of samples from each class in each source domain. First, the number of samples in each source domain dataset is counted according to class, and then divided by the total number of samples in the corresponding class to obtain the weight of the classifier for each class. Based on these weights, the n predictions from the domain-specific branch are fused into a single prediction. Then, it is fused with the predicted value z0 of the domain-invariant branch using a hyperparameter α: After fusion by the hybrid module, the final predicted value is obtained.
[0089] To better explain the prior-based weight allocation strategy, please refer to... Figure 3This is an example of a priori-based weight allocation strategy. Suppose we are dealing with a four-class classification problem and have three source datasets available. To obtain the weights, we first count the number of samples in each dataset by class, then divide by the total number of samples in that class to get the percentage. We then concatenate the percentage values for all classes in a dataset into a vector; this vector represents the weights of the corresponding classifier in the domain-specific branch.
[0090] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.
Claims
1. An ECG signal classification system based on deep learning neural networks, characterized in that, include: A common feature extraction module, which uses a first convolutional neural network to extract shallow features from the input ECG signal; The domain-invariant branch module employs a second convolutional neural network to extract features from the shallow features. By minimizing the classification loss, it implicitly performs feature matching for all source domains to obtain a label prediction. The domain-specific branch module employs a third convolutional neural network and uses LMMD loss to perform feature matching on the input shallow features between each pair of source and target domains, thereby obtaining n label predictions. The hybrid module, based on prior knowledge, assigns different confidence levels to each sub-branch according to the number of samples of various labels that can be obtained from each sub-branch of the domain-invariant branch module and the domain-specific branch module. It then combines the n+1 label predictions obtained from the domain-invariant branch module and the domain-specific branch module to obtain the final label prediction. Specifically: The number of samples in each source domain dataset is counted according to the category, and then divided by the total number of samples in the corresponding category to obtain the weight of the corresponding classifier for each category sample. Based on the weights, the n predictions of the domain-specific branch are merged into a single prediction. The predicted value of the domain-invariant branch With a hyperparameter Fusion: The final predicted value was obtained. .
2. The ECG signal classification system based on a deep learning neural network according to claim 1, characterized in that, The first convolutional neural network of the common feature extraction module includes: A convolutional layer, comprising a normalization layer and a nonlinear layer, wherein the kernel size of the convolutional layer is 1. 7; The number of max pooling layers is two. A residual module, which is located after the max pooling layer; The input to the first convolutional neural network is all ECG signals, and the output is the shallow features of all ECG signal data.
3. The ECG signal classification system based on a deep learning neural network according to claim 1, characterized in that, The second convolutional neural network of the domain-invariant branch module includes a domain-invariant feature extractor and a classifier; The domain-invariant feature extractor includes a residual module, two max pooling layers, three residual modules with channel attention, and a global average pooling layer. The classifier includes a fully connected layer and a Sigmoid layer; The input to the second convolutional neural network is the shallow features of the ECG signal, and the output is a prediction of the classification result of the ECG signal data.
4. The ECG signal classification system based on a deep learning neural network according to claim 3, characterized in that, The channel attention includes: An average pooling layer pools the features of each channel into a single element; Two fully connected layers, each followed by a non-linear layer, and a final sigmoid layer, map the pooled features to the weights of the corresponding channels.
5. The ECG signal classification system based on a deep learning neural network according to claim 1, characterized in that, The third convolutional neural network of the domain-specific branch module includes: n sub-networks, the structure of which is the same as that of the second convolutional neural network, where n is the number of source domain datasets; The input to each sub-network is a pair of shared shallow features from the source and target domain datasets, which are matched using a domain adaptation loss; the output is a prediction of the classification result of the ECG signal data.
6. The ECG signal classification system based on a deep learning neural network according to claim 5, characterized in that, The hybrid module combines the n+1 label predictions output by the second and third convolutional neural networks based on prior statistical information of various labeled samples in the source domain dataset. ,in, {zi} represents the predicted values of the n branches of the domain-specific branch; z0 represents the predicted values of the domain-invariant branch; α∈[0,1] is the fusion factor.
7. A method for classifying ECG signals based on deep learning neural networks, characterized in that... , Obtain the source domain dataset and target domain dataset of the ECG signal; Based on the source domain dataset and the target domain dataset, a common feature extraction network, a domain-invariant branch network, and a domain-specific branch network are trained, specifically as follows: The ECG signals in the source domain dataset are input into a common feature extraction network to extract shallow features of the ECG signals; The shallow features are respectively input into the domain-invariant branch network and the domain-specific branch network to obtain one classification and n classifications; Calculate the two classification losses for the one category and the n categories; The ECG signals in the target domain dataset are input to the domain-specific branch to extract target domain features. Calculate the domain adaptation loss between the target domain features and the source domain features; Based on the domain adaptation loss and classification loss, the network parameters of the common feature extraction network, the domain invariant branch network and the domain-specific branch network are updated using the backpropagation method. The ECG signals from the target domain dataset are input into a pre-trained common feature extraction network, a domain-invariant branch network, and a domain-specific branch network. The domain-invariant branch will then provide a predicted value. Domain-specific branches will provide Predicted values A total of n+1 predicted categories were obtained; The n+1 predicted classifications are fused to obtain the ECG signal classification result, specifically: The number of samples in each source domain dataset is counted according to the category, and then divided by the total number of samples in the corresponding category to obtain the weight of the corresponding classifier for each category sample. Based on the weights, the n predictions of the domain-specific branch are merged into a single prediction. The predicted value of the domain-invariant branch With a hyperparameter Fusion: The final predicted value was obtained. .
8. The ECG signal classification method based on deep learning neural networks according to claim 7, characterized in that... The classification loss function is Focal loss. ; in, The predicted value of the model. This represents the true label of the sample. It is a hyperparameter that controls the penalty strength of the loss function; The domain adaptation loss is the LMMD loss. ； Where C represents the total number of categories, and c represents the category number. Represents the i-th source domain. Indicates the target domain. It comes from The sample, It comes from The sample, express The weight on the c-th label express The weight on the c-th label Represents the kernel function. This represents the regenerated Hilbert space.