An electrocardiogram classification method based on improved SSA and DMSANet

Abstract

The present application relates to a kind of electrocardiogram classification method based on improved SSA and DMSANet, acquires electrocardiogram signal, with improved SSA algorithm noise reduction processing;DMSANet is constructed, including backbone network, different scale sub-processing module and multiscale feature fusion module arranged in parallel, to DMSANet after the electrocardiogram signal of noise reduction processing classification.The present application significantly improves the effect of noise reduction, effectively improves the accuracy and signal quality of ECG signal;Based on the backbone network of ResNet architecture in DMSANet, features are learned and extracted from ECG signal, and the multi-subnet and multiscale feature fusion strategy help to capture and integrate feature information of different scales, improve the classification performance and robustness of ECG signal;Combined with improved SSA algorithm and DMSANet architecture, improve the overall reliability and application effect of classification research based on ECG signal, especially suitable for biomedical signal processing field.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing or generation in general, and in particular to an electrocardiogram classification method based on improved SSA and DMSANet in the field of signal processing and deep learning. BACKGROUND

[0002] Cardiovascular Disease (CVD) is a major challenge faced by people worldwide, which includes a variety of heart and blood vessel related diseases such as coronary heart disease, cerebrovascular disease, rheumatic heart disease and other cardiovascular diseases. According to statistics, about 17.9 million people die of cardiovascular disease every year, making it one of the leading causes of death from diseases worldwide.

[0003] Early diagnosis and intervention are crucial to reducing the mortality rate of cardiovascular disease. In terms of diagnosis, Electrocardiogram (ECG) is a major diagnostic tool that records the electrical signals generated by the heart with each beat, which can reflect the rhythm and functional status of the heart. By analyzing ECG, doctors can detect abnormal rhythms of the heart, myocardial ischemia, myocardial infarction signs and other pathological changes, thereby helping to diagnose and treat cardiovascular disease.

[0004] However, in the process of collecting electrocardiogram signals, they are often disturbed by various noises, including but not limited to motion artifacts, baseline drift, Electromyogram (EMG) interference, static interference, power line interference (50Hz) and respiration, etc. These interference factors seriously affect the clarity and accuracy of the electrocardiogram signals, thereby reducing the practical value of electrocardiogram in clinical diagnosis; in addition, analyzing electrocardiogram records is usually time-consuming and prone to human error, which can lead to misdiagnosis and only achieve minimal accuracy.

[0005] Therefore, it is urgent to develop an ECG filtering and classification algorithm that can effectively suppress various noise interference, ensure signal accuracy and integrity, and achieve effective signal classification processing. SUMMARY

[0006] To solve the above technical problems, the present application provides an electrocardiogram classification method based on improved SSA and DMSANet, which effectively processes the noise and classification problems faced by ECG.

[0007] The technical scheme adopted by the present application is an electrocardiogram classification method based on improved SSA and DMSANet, the method acquires an electrocardiogram signal, and the electrocardiogram signal is denoised by using an improved SSA algorithm; a DMSANet is constructed, including a backbone network, different-scale sub-processing modules arranged side by side, and a multi-scale feature fusion module, so that the DMSANet acquires classification of the denoised electrocardiogram signal.

[0008] Preferably, the improved SSA algorithm comprises a signal decomposition method, a pivot confirmation method and a signal reconstruction method arranged in sequence, the effective part of the electrocardiogram signal is acquired by using the pivot confirmation method after the electrocardiogram signal is decomposed by using the signal decomposition method, and the electrocardiogram signal is reconstructed by using the signal reconstruction method.

[0009] Preferably, the signal decomposition method comprises the following steps:

[0010] S1.1.1 let the electrocardiogram signal be a length sequence with a length of N, map it into a vector sequence with a length of L by using embedding operation, map it into an L×K space to obtain a trajectory matrix X, and 1<L<N;

[0011] S1.1.2 perform singular value decomposition on the trajectory matrix X.

[0012] Preferably, the pivot confirmation method comprises the following steps:

[0013] S1.2.1 construct a new index, and reconstruct the matrix of the electrocardiogram signal based on the new index;

[0014] S1.2.2 map the reconstructed matrix into a group of reconstructed one-dimensional time sequences by using a diagonal average method, and define each one-dimensional time sequence as a pivot of the original electrocardiogram signal;

[0015] S1.2.3 judge and screen effective pivots;

[0016] S1.2.4 extract features based on the effective pivots.

[0017] Preferably, S1.2.3 comprises the following steps:

[0018] S1.2.3.1 use any pivot as sample data, and use whether it is selected as a target pivot as a sample label, the sample label being 0 or 1;

[0019] S1.2.3.2 apply a logistic regression model to the sample data, combine a Sigmod function and a linear regression function, and calculate the probability that any sample data belongs to the sample label;

[0020] S1.2.3.3. Extracting key features from each single principal component based on a gradient descent algorithm and a cross-entropy loss function, taking the output of the logistic regression model as the input of the logistic regression model to determine whether the current principal component is an effective principal component.

[0021] Preferably, the dimensions of the features include frequency domain, nonlinearity, and time domain. The features extracted from the principal components here include but are not limited to signal mean value, signal mean square root, signal kurtosis value, etc., which can be determined as needed.

[0022] Preferably, the backbone network comprises a CNN backbone network, a convolutional layer, and two max-pooling layers arranged in sequence, and between the two max-pooling layers, three groups of first convolutional layers and CBAM modules are arranged in sequence, and four groups of second convolutional layers and CBAM modules are arranged in sequence.

[0023] Preferably, the DMSANet comprises two sub-processing modules; any sub-processing module comprises three groups of third convolutional layers and CBAM modules arranged in sequence, two groups of fourth convolutional layers and CBAM modules, and a global pooling layer.

[0024] Preferably, the loss optimization is based on the focal loss.

[0025] The application provides an electrocardiogram classification method based on improved SSA and DMSANet, acquires an electrocardiogram signal, and performs denoising processing on the electrocardiogram signal by using an improved SSA algorithm; constructs a DMSANet, including a backbone network, different scale sub-processing modules arranged side by side, and a multi-scale feature fusion module, and classifies the denoised electrocardiogram signal by using the DMSANet.

[0026] The application has the following beneficial effects:

[0027] (1) When processing a complete ECG signal, the improved SSA algorithm is used for denoising, and a logistic regression algorithm is introduced in the principal component reorganization stage to improve the principal component reorganization method to automatic reorganization, which significantly improves the denoising effect and effectively improves the accuracy and signal quality of the ECG signal;

[0028] (2) The backbone network based on the ResNet architecture in the DMSANet learns and extracts features from the ECG signal, and the multi-subnet and multi-scale feature fusion strategy helps to capture and integrate feature information of different scales, thereby improving the classification performance and robustness of the ECG signal;

[0029] (3) The improved SSA algorithm and the DMSANet architecture are combined to improve the overall reliability and application effect of the classification research based on the ECG signal, and are particularly suitable for the field of biomedical signal processing. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A method flowchart of the present application;

[0031] Figure 2 An operation flowchart between the initial electrocardiogram signal and the denoising electrocardiogram signal in the present application;

[0032] Figure 3 A structural schematic block diagram of the DMSANet in the present application;

[0033] Figure 4 A network structure parameter diagram corresponding to Figure 3 ;

[0034] Figure 5 A structural schematic diagram of the CBAM module in the present application;

[0035] Figure 6 An F1 score using the method of the present application. DETAILED DESCRIPTION

[0036] The present application will be further described in detail below in conjunction with examples, but the scope of protection of the present application is not limited thereto.

[0037] As shown in Figure 1 , the present application relates to an electrocardiogram classification method based on improved SSA and DMSANet, the method acquires an electrocardiogram signal, and the electrocardiogram signal is denoised by an improved SSA algorithm; a DMSANet is constructed, including a backbone network, different scale sub-processing modules arranged side by side, and a multi-scale feature fusion module, so as to classify the denoised electrocardiogram signal by the DMSANet.

[0038] The method will be described in detail below.

[0039] As shown in Figure 2 , the improved SSA algorithm includes a signal decomposition method, a pivot confirmation method and a signal reconstruction method arranged in sequence, the effective part of the electrocardiogram signal is obtained by the pivot confirmation method after the electrocardiogram signal is decomposed by the signal decomposition method, and the electrocardiogram signal is reconstructed by the signal reconstruction method.

[0040] In the present application, the ECG signal is denoised by the improved SSA algorithm, and the initial SSA algorithm is composed of signal decomposition and signal reconstruction;

[0041] In the signal decomposition process, the initial ECG signal is decomposed by embedding operation and singular value decomposition, realizing the mapping of the signal in the high-dimensional space in the independent dimension, so that the effective component and the noise component in the signal are separated;

[0042] Then, by cooperating with the single pivot diagonal average, feature extraction and logistic regression algorithm, the corresponding effective component is selected in the high-dimensional space, and the noise component is removed.

[0043] In the signal reconstruction stage, the target signal is reconstructed, and the obtained is the ECG signal after noise reduction.

[0044] Specifically, the signal decomposition method comprises the following steps:

[0045] S1.1.1 Let the electrocardiogram signal be a length sequence with a length of N, map it into a vector sequence with a length of L by embedding operation, map it into an L×K space, and obtain a trajectory matrix X, 1<L<N;

[0046] Here, x=(x1, …, xN) is the research object, where N>2, and x is a non-zero sequence, that is, there is at least an integer i between 1 and N, such that x N ≠0; i

[0047] Let an integer L (1<L<N) be the window length, and K=N-L+1, the embedding operation maps the original time sequence (electrocardiogram signal sequence) into a vector sequence with a length of L, and maps it into an L×K space, to obtain a trajectory matrix X, which satisfies

[0048]

[0049] S1.1.2 Singular value decomposition is performed on the trajectory matrix X;

[0050] In the present application, singular value decomposition is performed on the trajectory matrix, and the trajectory matrix is expressed as a sum of first-order biorthogonal matrices;

[0051] Let the covariance matrix S=XX T , λ1, λ2…λ L be eigenvalues of S and λ1≥…≥λ L ≥0, and the singular value decomposition of X satisfies,

[0052]

[0053] Where d=rank(X), U and V are left and right singular matrices of the trajectory matrix X respectively;

[0054] Here, λ1, λ2…λ L are eigenvalues of the covariance matrix, and are the numerical values of the intermediate diagonal matrix in the following decomposition.

[0055] In fact, the signal reconstruction stage comprises two steps of grouping and diagonal line averaging, in the grouping stage, the pivot sequence of the index set {1, 2, …, d} can be divided into several discontinuous subsets according to actual needs, and the purpose of grouping is to find components suitable for electrocardiogram reconstruction, and here the pivot confirmation method is realized;

[0056] ​Specifically, the principal component confirmation method comprises the following steps:

[0057] S1.2.1 constructing a new index, reconstructing the matrix of electrocardiogram signals based on the new index, i.e. selecting corresponding effective components in a high-dimensional space to recombine and eliminate noise components, thereby reconstructing the target signal to achieve the purpose of noise reduction;

[0058] Let I = {i1, … i p}(p < d) be the corresponding new index, then the matrix corresponding to the ECG can be represented as,

[0059]

[0060] S1.2.2 mapping the reconstructed matrix X group to a set of reconstructed one-dimensional time series x recon using the diagonal average method, represented as

[0061]

[0062] Where A8 = {(l, k): l + k = s + 1, 1 ≤ l ≤ L, 1 ≤ k ≤ K}, l and k correspond to L and K in X group , represent the variables of the index, and N(A s ) is the number of elements in set A s ;

[0063] After reconstructing the diagonal average, the original signal is given, with x as the filtered electrocardiogram signal, and the original one-dimensional time series signal x can be represented as:

[0064] x = x recon(1) +x recon(2) +…+x recon(d)

[0065] Where x recon(n) is reconstructed from the grouping matrix X group(n) (n = 1, …, d), and x recon(n) is defined as the principal component of the original electrocardiogram signal; the original signal is recovered by superimposing multiple principal components, and each x recon(n) here is one of the important components of the signal.

[0066] S1.2.3 judging and screening effective principal components;

[0067] The selection of principal component sequences is crucial in the signal reconstruction stage of singular spectrum analysis. However, the selection of principal component sequences usually relies on the experience of researchers and lacks universal standards. In addition, for different types of signals, the positions of principal components of effective components in the sequence are not fixed, which makes it difficult to achieve effective noise reduction using principal components with fixed positions and also limits the automated execution of singular spectrum analysis algorithms. Based on this, a logistic regression algorithm is proposed. By constructing feature engineering based on ECG signals and noise signals, the algorithm judges each principal component obtained after signal decomposition to determine which principal components constitute the effective components of the ECG signal, thereby achieving automated selection of principal components, reducing manual intervention, and improving the automation level of singular spectrum analysis algorithms.

[0068] Specifically, S1.2.3 includes the following steps:

[0069] S1.2.3.1 Using any principal element x recon(i) As sample data, the corresponding label is whether it is selected as the target principal component, and the sample label is 0 or 1. i ∈{0,1}; each pair of sample data and sample label forms a sample class pair (x recon(i) y i );

[0070] S1.2.3.2 Apply the logistic regression model to the sample data, combining the Sigmoid function g(z) = 1 / (1+ze) and the linear regression function f(x) = ω t Combining x and b, calculate the probability that any sample data belongs to the sample label, satisfying...

[0071]

[0072] The parameters weight ω and intercept b are estimated using the maximum likelihood method, and the cross-entropy loss function can be obtained after transformation.

[0073]

[0074] S1.2.3.3 uses the gradient descent algorithm based on the cross-entropy loss function. Specifically, based on the cross-entropy loss function, the optimal solution of weight ω and intercept b is obtained by minimizing the difference between prediction and reality and minimizing the loss function. Key features are extracted from each single principal component and used as input to the logistic regression model to improve the classification performance of the model.

[0075] Considering model complexity, features are extracted from single principal component signals according to dimensions such as frequency domain, nonlinearity, and time domain. In the principal component selection step of singular spectrum analysis, the output of the logistic regression model is used to determine whether the current principal component is a valid principal component, thereby realizing automatic principal component selection.

[0076] S1.2.4 Based on the effective principal component feature extraction, a certain number of features are extracted from the effective principal component signal, including but not limited to the signal mean, signal kurtosis, frequency of the principal energy spectrum, and sample entropy of the signal.

[0077] like Figure 3 , Figure 4 As shown, in the specific implementation process, a deep multi-scale attention network (DMSANet) is used for ECG classification. First, the backbone network is used for learning and feature extraction to obtain the original feature representation. The features are fed into two sub-processing modules to extract features at different scales. Each sub-processing module focuses on processing feature maps at a specific scale to capture more detailed local features and broader global features. Subsequently, after feature extraction at different scales, these feature maps are fused together.

[0078] The backbone network includes a CNN backbone network, convolutional layers and two max pooling layers arranged in sequence. Between the two max pooling layers, there are three combinations of first convolutional layers and CBAM modules and four combinations of second convolutional layers and CBAM modules arranged in sequence.

[0079] In this invention, during specific implementation, the number of channels in the convolutional layer and the first convolutional layer is 64, and the number of channels in the second convolutional layer is 128, in order to capture more features.

[0080] Specifically, ResNet has excellent performance, making it ideal for building backbone networks; such as Figure 5 As shown, CBAM, as an attention mechanism module for enhancing the performance of convolutional neural networks (CNNs), improves the model's perceptual ability by introducing channel attention and spatial attention into CNNs, thereby improving performance without increasing network complexity.

[0081] The LeakyReLU activation function addresses the "dead neuron" problem of the traditional ReLU function by providing small gradients in the negative region, maintaining gradient flow in the network, and improving the stability of the training process.

[0082] The DMSANet includes two sub-processing modules; each sub-processing module includes a combination of three sets of third convolutional layers and CBAM modules arranged in sequence, a combination of two sets of fourth convolutional layers and CBAM modules, and a global pooling layer.

[0083] In this step, the model's feature maps are passed to two sub-processing modules, which extract features at different scales. Each sub-network focuses on processing feature maps at a specific scale to obtain more refined local features and more macroscopic global features. By extracting and processing features at different scales, the model can more comprehensively understand the multi-level information in the data, thereby improving its ability to recognize complex patterns and thus improving the accuracy of classification or detection.

[0084] In this invention, the third convolutional layer has 256 channels and the fourth convolutional layer has 512 channels, which are used to capture more features at different scales.

[0085] After feature extraction, the feature maps obtained from different sub-processing modules are fused to integrate features at various scales, so as to optimize the loss more effectively.

[0086] This invention uses focus loss for loss optimization to emphasize difficult-to-classify samples and reduce the impact of easily-classified samples on loss calculation.

[0087] Focal Loss(p tj )=-α t (1-p tj ) γ logp tj

[0088] Where, p tj It is the predicted probability of the model, α t It is a balancing factor used to balance the weights of positive and negative classes. γ is an adjustment factor, usually a positive number, used to adjust the weights of samples that are difficult to classify.

[0089] In the ECG dataset, class imbalance is common, with common diseases usually far outnumbering uncommon diseases. Therefore, the network needs to focus not only on the most common diseases but also on the less common ones. Focal Loss can be used to address class imbalance by reducing the weight of easily classified samples and increasing the weight of difficult-to-classify samples. This allows the model to focus more on challenging samples, thereby improving the model's performance on imbalanced datasets.

[0090] The following is a specific example.

[0091] The PTB-XL ECG dataset is a large dataset containing 21,837 clinical 12-lead electrocardiograms from 18,885 patients, each recording lasting 10 seconds. The raw waveform data was annotated by up to two cardiologists, who assigned multiple possible ECG statements to each record. The dataset includes 71 ECG statements conforming to the SCP-ECG criteria, covering diagnostic, morphological, and rhythmic information.

[0092] For detailed data on PTB-XL, please refer to Tables 1 and 2;

[0093] Table 1

[0094]

[0095] Table 1 shows the six diagnostic categories in the dataset and the corresponding number of ECG signals;

[0096] Table 2

[0097]

[0098] Table 2 lists the superdiagnoses and subdiagnoses;

[0099] The dataset covers a variety of clinical diseases and has a high degree of credibility.

[0100] The network is trained using a dataset, and its performance is measured by F1 score, recall, and precision, as shown in the following formulas:

[0101]

[0102] In this model, TP represents true positive, indicating that the model correctly predicted a positive instance as positive; FN represents false negative, indicating that the model incorrectly predicted a positive instance as negative; FP represents false positive, indicating that the model incorrectly predicted a negative instance as positive. Recall can be seen as not missing any people with heart disease, while Precision can be seen as avoiding misdiagnosing healthy people as having heart disease. F1 represents the balance between recall and precision. In the ECG dataset, the F1 score is chosen as the evaluation metric. Figure 6 The network's F1 scores were shown for 15 different diseases, with F1 scores exceeding 0.8 for 10 of them.

[0103] In summary, this example uses an improved SSA algorithm to filter the signal and DMSANet for classification. This not only improves the quality of the ECG signal but also enhances the accuracy of the classification.

[0104] The present invention also relates to a computer-readable storage medium storing an electrocardiogram (ECG) classification program based on improved SSA and DMSANet, which, when executed by a processor, implements the aforementioned ECG classification method based on improved SSA and DMSANet.

[0105] The present invention also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the above-mentioned electrocardiogram classification method based on the improved SSA and DMSANet.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0111] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An electrocardiogram classification method based on improved SSA and DMSANet, characterized in that: The method acquires electrocardiogram (ECG) signals and uses an improved SSA algorithm to denoise the ECG signals. The improved SSA algorithm includes a signal decomposition method, a principal component confirmation method, and a signal reconstruction method set in sequence. The principal component confirmation method includes the following steps: S1.2.1 Construct a new index and reconstruct the matrix of the electrocardiogram signal based on the new index; S1.2.2 The reconstructed matrix is ​​mapped to a set of reconstructed one-dimensional time series using the diagonal averaging method, and each one-dimensional time series is defined as the principal component of the original electrocardiogram signal; S1.2.3 Identify and filter valid principal components; S1.2.4 Feature extraction based on effective principal components; DMSANet is constructed, including a backbone network, parallel sub-processing modules of different scales, and a multi-scale feature fusion module, to classify the denoised electrocardiogram signals.

2. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: After decomposing the electrocardiogram (ECG) signal using the signal decomposition method, the effective part of the ECG signal is obtained using the principal component confirmation method, and the ECG signal is reconstructed using the signal reconstruction method.

3. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 2, characterized in that: The signal decomposition method includes the following steps: S1.1.1 Let the electrocardiogram signal be a sequence of length N. Using embedding operations, map it into a vector sequence of length L, mapping it into an L×K space to obtain the trajectory matrix X. <L<N; S1.1.2 Perform singular value decomposition on the trajectory matrix X.

4. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: S1.2.3 includes the following steps: S1.2.3.1 Using any principal element as sample data, the corresponding sample label is whether it is selected as the target principal element, and the sample label is 0 or 1; S1.2.3.2 The logistic regression model is applied to the sample data, and the Sigmoid function and the linear regression function are combined to calculate the probability that any sample data belongs to the sample label; S1.2.3.3 uses the gradient descent algorithm, based on the cross-entropy loss function, to extract key features from each single principal component, which are used as input to the logistic regression model. The output of the logistic regression model is used to determine whether the current principal component is a valid principal component.

5. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: The dimensions of the features include frequency domain, nonlinearity, and time domain.

6. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: The backbone network is based on ResNet and includes a combination of two sets of convolutional layers and CBAM modules, using LeakyReLU for nonlinear transformation and feature extraction.

7. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 6, characterized in that: The backbone network includes a CNN backbone network, convolutional layers and two max pooling layers arranged in sequence. Between the two max pooling layers, there are three combinations of first convolutional layers and CBAM modules and four combinations of second convolutional layers and CBAM modules arranged in sequence.

8. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: The DMSANet includes two sub-processing modules; each sub-processing module includes a combination of three sets of third convolutional layers and CBAM modules arranged in sequence, a combination of two sets of fourth convolutional layers and CBAM modules, and a global pooling layer.

9. The electrocardiogram classification method based on improved SSA and DMSANet according to claim 1, characterized in that: Loss optimization is performed based on focus loss.

Images