Electrocardiogram classification method based on multi-modal feature fusion and fuzzy integral decision

Abstract

The application provides a kind of electrocardiogram classification method based on multi-modal feature fusion and fuzzy integral decision, belongs to artificial intelligence and medical technology field, solved the technical problem of insufficient multi-modal fusion, low decision reliability in prior art.The technical scheme is: through discrete wavelet transform extraction time domain and frequency domain statistical characteristics, and utilize the deep encoder based on Transformer to extract cross lead time-space feature;Adopt bidirectional cross attention mechanism, capture intermodal complementarity;Then build three types of classifiers and introduce fuzzy integral to make decision fusion, through fuzzy measure modeling interdependence between classifiers, dynamically allocate weight.The beneficial effects of the application are: the application combines clinical interpretable features with data-driven models, improves multi-modal feature fusion depth through bidirectional attention mechanism, enhances decision robustness using fuzzy integral, significantly improves the accuracy and interpretability of automatic diagnosis of cardiovascular diseases.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical health technology, specifically to an electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision. Background Technology

[0002] Cardiovascular disease is one of the leading causes of death worldwide, and its early and accurate diagnosis is crucial for reducing patient mortality. Electrocardiography (ECG), as a non-invasive, economical, and widely used diagnostic tool, plays an irreplaceable role in the clinical diagnosis of cardiovascular diseases. However, traditional ECG analysis relies heavily on physicians' subjective experience, resulting in low diagnostic efficiency and problems such as inconsistent diagnostic criteria and poor reproducibility. With the development of artificial intelligence technology, automated ECG diagnostic methods based on machine learning and deep learning have gradually become a research hotspot, but existing technologies still face several key challenges that urgently need to be addressed.

[0003] In feature extraction, existing technologies mainly fall into two categories: one is based on traditional machine learning methods, such as Support Vector Machines (SVM) and Principal Component Analysis (PCA). These methods require manual feature design, relying not only on expert domain knowledge but also exhibiting poor robustness against noise interference or signal distortion. The other is based on deep learning methods, such as Convolutional Neural Networks (CNN) and Transformer models. While these can automatically extract features, the extracted features lack clinical interpretability and are difficult to correlate with doctors' diagnostic logic. Regarding multimodal feature fusion, existing technologies generally employ simple feature concatenation or weighted averaging strategies, failing to fully consider the nonlinear interactions between different feature modalities. Furthermore, existing methods often focus only on the fusion of high-level abstract features, neglecting important diagnostic information contained in low-level features, resulting in information loss. In decision fusion, existing ECG classification systems mostly employ independent classifiers or majority voting strategies, failing to fully consider the dependencies between classifiers. Traditional weighted averaging methods, unable to model the collaborative or competitive relationships between classifiers, exhibit significant performance degradation when faced with imbalanced data or noise interference. Summary of the Invention

[0004] This invention provides an electrocardiogram (ECG) classification method based on multimodal feature fusion and fuzzy integral decision-making. This method aims to solve the technical problems of insufficient multimodal fusion and insufficient decision reliability in existing automatic ECG classification technologies.

[0005] The main inventive idea of ​​this invention is as follows: The core of this invention lies in constructing an electrocardiogram (ECG) classification method that integrates clinical knowledge guidance, dynamic interaction of multimodal features, and fuzzy integral decision-making. This method first extracts clinically significant statistical features from ECG signals through a knowledge-guided feature extraction module, and then extracts deep features representing signal patterns and structures through a neural network. Next, it utilizes a bidirectional cross-attention mechanism to achieve deep interaction and fusion of different modal features. Finally, it employs a fuzzy integral-based decision-making method, fully considering the dependencies between classifiers, significantly improving the accuracy and reliability of diagnostic results.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: an electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision, comprising the following steps:

[0007] S1. The system preprocesses and analyzes the raw multi-lead electrocardiogram signals, extracting a set of interpretable and discriminative statistical features based on clinical medical knowledge. This module utilizes discrete wavelet decomposition and reconstruction to comprehensively reflect the dynamic changes in cardiac electrical activity, providing interpretable input information for subsequent diagnostic models.

[0008] S2. The system utilizes neural networks to perform end-to-end modeling of the raw electrocardiogram (ECG) signal. By constructing a time-series segmentation structure based on leads and introducing position encoding and lead embedding, the signal is transformed into a unified multidimensional representation. Subsequently, this input is fed into a Transformer-based deep encoder to extract spatiotemporal features across leads and time slices, outputting deep feature vectors representing signal patterns and structures. This module can automatically learn pathological information hidden in time-series data, improving the model's recognition capabilities.

[0009] S3. The system constructs a bidirectional cross-attention mechanism, mapping statistical features and deep coding features to the same feature space, and performing bidirectional attention interaction to capture the correlation and complementarity between modalities. The fusion module outputs a joint feature representation, providing a more comprehensive diagnostic basis for subsequent decision-making.

[0010] S4. The system constructs classifiers based on statistical features, deep features, and fusion features, respectively, and outputs multiple prediction results. By introducing Choquet fuzzy integrals, the outputs of multiple classifiers are weighted and fused using fuzzy measures, thereby enhancing the system's stability and decision reliability.

[0011] S5. The system integrates the above modules into an end-to-end diagnostic process, trains and validates using a standard electrocardiogram dataset, adopts a joint optimization strategy to learn the model parameters and fuzzy measures simultaneously, and finally outputs the target disease category and performs performance evaluation.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S11. Acquire standard 12-lead electrocardiogram (ECG) signals. Each lead signal has a length of T, and there are a total of L = 12 leads, forming the original input matrix. Where N is the number of samples;

[0014] S12. Perform uniform preprocessing on all lead signals using a 4th-order Butterworth bandpass filter with a frequency band of 0.67–48 Hz to remove high-frequency noise, electromyographic interference, and low-frequency baseline drift. Then, perform normalization to unify the signal amplitude to the range [-1, 1], thus obtaining the signal input matrix.

[0015] S13. The second lead signal is decomposed into 7 levels using discrete wavelet transform. The Daubechies wavelet is used as the mother wavelet function to extract 7 sets of detail coefficients (D1–D7) and 1 set of approximation coefficients (A7) to obtain multi-resolution information of the signal.

[0016] S14. Calculate the following statistical characteristics in the original signal and the reconstructed signal respectively:

[0017] Temporal characteristics: RR interval, PR interval, QRS duration, QT interval, heart rate (HR), heart rate variability (SDNN, RMSSD), etc.

[0018] Frequency domain characteristics: power spectral entropy, Shannon entropy, sample entropy, signal complexity, kurtosis, and skewness, etc.

[0019] S15. Stack the D1-dimensional statistical features extracted from the second lead in order to form a statistical feature matrix. The Z-score standardization method was used to unify its dimensions, so that each feature dimension has a distribution with a mean of 0 and a standard deviation of 1.

[0020] S16, F stat As the output of the statistical feature extraction module, it is used by the subsequent fusion module and classifier, and the lead dimension information is retained for subsequent feature space matching.

[0021] Furthermore, step S2 specifically includes the following steps:

[0022] S21. The preprocessed signal from step S1 is... For the signal of each lead Divide the data into N patches, each patch having a length of p.

[0023] S22. Project each patch onto a fixed dimension d using a fully connected layer. patch This forms an embedded feature representation;

[0024] S23. Design two embedding mechanisms to enhance input structure information:

[0025] Position embedding Learnable vectors are used to distinguish the temporal location of patches in lead signals. Lead embedding Assign a unique, learnable vector to each lead to represent its spatial location;

[0026] S24. Concatenate all lead patches into an input tensor. The input is fed into the Transformer encoder module to construct a multi-layer encoder for modeling intra-patch temporal dependencies and inter-patch cross-lead spatial correlations, where d emb The dimensions to be embedded after splicing;

[0027] S25. After encoding, the output vectors of each lead are aggregated using average pooling to form the representation vector of each sample. As the output of the deep modeling module, D2 is the dimension of the representation vector;

[0028] Furthermore, step S3 specifically includes the following steps:

[0029] S31, Statistical characteristics F stat With depth features F ecg Each is mapped to a unified dimension d through a linear transformation. fusion Retrieve the corresponding query, key and value:

[0030] Q1, K1, V1 = F stat W1

[0031] Q2, K2, V2 = F ecg W2

[0032] Where Q1, K1, and V1 are the queries, keys, and values ​​obtained by linear mapping of statistical features, respectively, and W1 is the learnable matrix of linear mapping of statistical features; Q2, K2, and V2 are the queries, keys, and values ​​obtained by linear mapping of deep features, respectively, and W2 is the learnable matrix of linear mapping of deep features.

[0033] S32. Calculate the depth feature F of the ECG signal. ecg Relative to statistical characteristic F stat Attention is calculated by pairing the query Q1 from the statistical features with the keys K2 and V2 from the deep features, thus aligning the attention of the statistical features with the deep features, resulting in the intermediate representation F. 12 :

[0034]

[0035] S33, Reverse calculation of ECG signal statistical characteristics F stat Relative to depth features F ecg Attention is calculated by aligning the query Q2 from the deep features with the keys K1 and V1 from the statistical features, thus achieving attention alignment guided by the statistical features for the deep features, resulting in the intermediate representation F. 21 :

[0036]

[0037] S34, F 12 With F 21 At the feature dimension concatenation stage, a learnable weight matrix W3 integrates the connected features into a unified fused feature representation F. fusion :

[0038] F fusion =Concat(F 12 ,F 21 W3

[0039] Concat(.) represents the matrix concatenation operation.

[0040] S35, Output F fusion Used for subsequent sub-model classification and fuzzy integration processes.

[0041] Furthermore, step S4 specifically includes the following steps:

[0042] S41, F stat F ecg and F fusion Input the data into the classifier and output the corresponding data. This represents the predicted probability for C disease categories;

[0043] S42, Transform the vector The components are concatenated to form the classifier output matrix.

[0044]

[0045] Matrix element y ij ∈Y, row index i∈{1,2,3} corresponds to the prediction results of statistical features, deep features and fusion features respectively, and column index j∈{1,2,...,C} represents the disease category number.

[0046] S43. Constructing a fuzzy measure: First, define a differentiable fuzzy measure function g. θ :2 {1,2,3} →[0,1], where indices 1, 2, and 3 correspond to the statistical feature prediction results, respectively. Prediction results of deep features and prediction results of fusion features The fuzzy measure function is dynamically modeled through the parameter vector θ=[θ1, θ2, ..., θ7], with a parameter dimension of 7, covering the measure values of all non-empty subsets. To satisfy the mathematical constraints of fuzzy measures, g θ (.) denotes the mapping from the set to the parameter θ i , and the following constraints can be adopted:

[0047] 1. Boundary condition constraint: By setting and g θ ({1,2,3})=1, the completeness of the measure space is ensured;

[0048] 2. Monotonicity constraint: A regularization term is introduced into the loss function to penalize parameter update directions that violate monotonicity, where λ is a hyperparameter.

[0049] S45. Fuzzy measure parameterization: First, define learnable parameters α1, α2, and α3, which are mapped to independent weights through a sigmoid function to quantify the independent decision contribution of each model. Define interaction parameters β 12 , β 21 and β 23 , and use the nonlinear function φ(x)=ln(1+e x ) to calculate the joint weight increment. The joint weight of model i and model j where i<j is defined as:

[0050] g({i,j})=min(θ i +θ j +φ(β ij ),1)

[0051] Then the weights are normalized to meet the constraint that the full set measure of the fuzzy measure is 1.

[0052] S46. To realize gradient descent optimization, a differentiable fuzzy integral calculation process is designed:

[0053] First, extract the classifier output matrix Y, and generate a matrix by sorting the column elements in descending order Then, based on the matrix S, calculate the Choquet integral weights:

[0054] w1=g θ ({σ1})

[0055] w2=g θ ({σ1,σ2})-g θ ({σ1})

[0056] w3=1-g θ ({σ1,σ2})

[0057] Where σ i This represents the model index corresponding to the i-th sorting position determined by matrix S.

[0058] Finally, the fusion prediction value for category j is calculated.

[0059]

[0060] S46. Calculate the prediction probability The cross-entropy loss function between the true label and the actual label is used to calculate the gradient and update the parameter θ through backpropagation.

[0061] Furthermore, step S5 specifically includes the following steps:

[0062] S51. Combine modules S1 to S4 into an end-to-end trainable network, with 12-lead ECG as input and target cardiovascular disease classification results as output.

[0063] S52. Use multiple public datasets (such as PTB-XL, Chapman, CPSC2018) for training and evaluation. Standardize the samples and divide them into training set, validation set and test set.

[0064] S53, The loss function is weighted cross-entropy, the optimizer is AdamW, the initial learning rate is set to 1e-4, the batch size is 64, the training is 100 rounds, and early stopping is used to prevent overfitting;

[0065] S54. In the testing phase, perform independent reasoning for each input sample, output the final classification result, and evaluate it in five dimensions: accuracy, sensitivity, specificity, F1 score, and AUROC value.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] 1. Enhanced deep interaction capability of multimodal feature fusion. Through a bidirectional cross-attention mechanism, the system can achieve dynamic interaction between statistical features (time-domain indicators, signal entropy, etc.) and deep spatiotemporal features (translead dependencies), solving the problem of insufficient utilization of intermodal information complementarity caused by feature splicing or weighted averaging in traditional methods, and enhancing the representational capability of fused features.

[0068] 2. Optimization of interpretability and robustness of diagnostic decisions. By combining knowledge-guided feature extraction (wavelet decomposition to reconstruct signals, heart rate variability indicators) with a dynamic fuzzy integral fusion strategy, the system retains the advantages of data-driven approaches while incorporating clinical diagnostic logic. The statistical feature matrix directly corresponds to commonly used diagnostic indicators by doctors (such as QT interval and heart rate variability), supporting the interpretability analysis of the decision-making process; fuzzy integrals suppress noise interference through dynamic weight allocation, improving robustness. Attached Figure Description

[0069] Figure 1 This is a block diagram of the electrocardiogram classification system using multimodal feature fusion and fuzzy integral decision-making in this invention;

[0070] Figure 2 This is a block diagram of signal decomposition and reconstruction based on discrete wavelet transform in this invention;

[0071] Figure 3 This is a feature fusion block diagram based on the bidirectional cross-attention mechanism in this invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] Example 1

[0074] This embodiment provides an electrocardiogram (ECG) classification method based on multimodal feature fusion and fuzzy integral decision, including the following steps:

[0075] S1. First, the acquisition and preprocessing of the raw multi-lead electrocardiogram signals are performed. This operation provides standardized and clinically meaningful input data for subsequent feature extraction and diagnostic models. The specific steps are as follows:

[0076] First, a standard 12-lead ECG signal was acquired and resampled to a frequency of 250Hz. For each sample, 10 seconds of continuous ECG activity data were recorded. To improve signal quality, the system performed uniform bandpass filtering and notch filtering on the raw ECG signal. The bandpass filter used a four-Butterworth filter with a passband range of 0.67Hz to 48Hz, effectively filtering out electromyographic interference and low-frequency baseline drift. To eliminate mains interference, the system further applied a 50Hz notch filter to each lead signal. Subsequently, the system performed artifact removal and normalization. For potential sudden spikes and baseline fluctuation artifacts, a combination of wavelet packet thresholding and moving average strategies was used to smooth the waveform contour and improve the signal stability of the ECG. To achieve consistent input amplitude across samples, the system normalized the maximum and minimum values ​​of each lead signal, scaling all signals uniformly to the [-1, 1] interval. The final signal input matrix was then obtained. Where L represents the number of leads, N represents the total number of samples, and T represents the number of sampling points per lead.

[0077] After the above preprocessing operations are completed, the system selects the second lead as the standard lead for statistical feature extraction. Based on commonly used clinical methods, the signal from this lead is subjected to a seven-level Daubechies wavelet decomposition, obtaining seven sets of detail coefficients D1, D2, ..., D7 and one set of approximation coefficients A7. After wavelet transform, the system performs inverse transform reconstruction on the signal at each scale, thereby separating the signal components of different frequency bands, laying the foundation for subsequent frequency domain and time domain feature analysis.

[0078] Next, the system extracts statistical features with clear clinical interpretability from the original signal and the wavelet reconstructed signal, including but not limited to: RR interval, PR interval, QRS width, QT interval, mean heart rate (HR), heart rate variability (SDNN, RMSSD), power spectral entropy, Shannon entropy, sample entropy, signal complexity, kurtosis, and skewness. All feature extraction is based on international electrocardiogram diagnostic standards and key indicators defined by clinicians, ensuring interpretability and medical consistency.

[0079] After statistical feature extraction, the system stacks the features in a predetermined order to form a feature matrix. The system calculates the mean and standard deviation for each dimension of the feature matrix column-by-column and performs Z-score standardization. The transformation formula is as follows:

[0080]

[0081] Where μ j σ j Let be the mean and standard deviation of the j-th feature in the training samples, respectively. and x ijThese are elements in the transformed and original feature matrices, respectively. The standardization process aims to eliminate the influence of different units on model training, improving model convergence speed and performance stability.

[0082] Through the above steps, the system completes the standardized preprocessing of the raw ECG signal and the extraction of statistical features with medical guidance significance, and outputs a standardized statistical feature matrix. With the preprocessed lead signal matrix This is intended for use by subsequent deep feature extraction and feature fusion modules.

[0083] S2. Deep Feature Learning and Modeling. This step aims to fully explore the nonlinear structural information, local waveform variability, and translead temporal patterns present in the electrocardiogram (ECG) signal. The operation process is as follows:

[0084] First, the system divides each preprocessed lead signal into several equal-length patches in chronological order, with each patch containing 250 sampling points. For a 10-second, 2500-point input signal, this process divides each lead into 20 non-overlapping patches. This segmentation strategy preserves the local dynamic characteristics of the signal and provides a temporal modeling basis for subsequent models.

[0085] For each segmented patch, the system first extracts its local contextual features using a one-dimensional convolution operation. The convolution kernel length is 3, the stride is 1, and the number of output channels is 64. These features are then fed into a fully connected layer for further mapping to a 64-dimensional embedding space.

[0086] To enhance the temporal sequence information and lead differentiation information, the system adds two learnable embedding vectors to each patch: position embedding and lead embedding. The position embedding represents the temporal position of the patch in the signal sequence, and the lead embedding identifies its corresponding ECG lead number. The sum of these three vectors yields the final embedding representation tensor.

[0087] Subsequently, the system inputs this multidimensional representation into a sequence modeling module consisting of a multi-layer Transformer encoder. The Transformer module contains a 6-layer stacked self-attention encoder, each layer containing 8 attention heads, with a hidden layer dimension of 256. It uses a combination of residual connections and feedforward network structures to achieve high-order nonlinear interactive modeling across patches and leads.

[0088] After acquiring the encoder output, the system performs average pooling on the output of each lead, thereby generating a set of deep temporal feature representations for each sample. This feature matrix is ​​denoted as...

[0089] Through the above operations, the system not only preserves the information of local ECG signal segments, but also models the contextual dependencies between different leads, providing important support for building a unified representation in the subsequent multimodal fusion stage.

[0090] S3. After extracting statistical and deep features, the system enters the multimodal feature fusion stage. The core objective of this stage is to unify and complement the feature vectors from two different modalities, thereby improving the discriminative ability of the downstream classifier. The fusion process is implemented using a bidirectional cross-attention mechanism and includes the following steps:

[0091] First, the system uses two linear mapping modules to convert the statistical feature matrix F into... stat and depth feature matrix F ecg Projecting onto a unified dimensional space yields intermediate feature representations. The dimensions of the linear mapping matrices W1 and W2 are 50×128 and 256×128, respectively.

[0092] Subsequently, the system constructs a bidirectional cross-attention structure: on one hand, statistical features are used as query terms, and deep features are used as key-value pairs; on the other hand, deep features are used as query terms, and statistical features are used as key-value pairs, as shown in the following expression:

[0093] Q1, K1, V1 = F stat W1

[0094] Q2, K2, V2 = F ecg W2

[0095] Where Q1, K1, and V1 are the queries, keys, and values ​​obtained by linear mapping of statistical features, respectively, and W1 is the learnable matrix of linear mapping of statistical features; Q2, K2, and V2 are the queries, keys, and values ​​obtained by linear mapping of deep features, respectively, and W2 is the learnable matrix of linear mapping of deep features.

[0096] Through the above calculations, the system obtains output vectors in two fusion directions:

[0097]

[0098] These two vectors respectively express the response characteristics of statistical information to the depth representation, and the response characteristics of the depth representation to the statistical features.

[0099] Next, the system will F 12 With F 21 A fusion matrix is ​​formed by concatenating features along the feature dimension, and then processed through a third linear mapping. Projecting onto the final fusion dimension yields the fusion feature matrix.

[0100] Through the aforementioned bidirectional attention mechanism and projection transformation, the system completes the unified modeling of multi-source modal information and outputs a feature matrix. It will be used as input for subsequent classifiers.

[0101] S4. In the fourth step of this invention, the system uses the three types of feature expressions obtained in the statistical feature extraction, deep feature encoding and multimodal fusion stages to construct three independent classification models, and introduces a differentiable Choquet fuzzy integral mechanism to dynamically weight and integrate their prediction outputs to enhance the discriminative power and robustness of the models.

[0102] First, the system establishes a two-layer feedforward neural network structure as a classifier for each type of feature, as follows: For statistical features (dimension 50), the network structure is 128→64→5, with ReLU activation function used between fully connected layers; for deep features (dimension 256), the network structure is 256→128→5; for fusion features (dimension 256), the network structure is 256→128→5; the five-dimensional vector output by the last layer represents the predicted probability of the corresponding five types of cardiovascular diseases, satisfying the normalization condition.

[0103] The system denotes the predicted vector output by each classifier as follows: To achieve the final decision output, the system introduces a decision function based on Choquet fuzzy integrals.

[0104] To further illustrate the specific operation of this fuzzy integral fusion strategy, a simplified numerical calculation example of a three-class classification task is given below. Suppose that this system performs a three-class classification task on a certain electrocardiogram sample, with the three categories being: Category 1 is Normal, Category 2 is Atrial Fibrillation, and Category 3 is Atrioventricular Block.

[0105] The prediction results of the system's three classifiers for this sample are shown in the table below:

[0106] Table 1. Prediction results of different classifiers for the samples.

[0107] Statistical Feature Classifier 0.20 0.60 0.30 Deep feature classifier 0.10 0.80 0.50 Fusion feature classifier 0.15 0.75 0.60

[0108] Taking category 2 as an example, let's perform fuzzy integral calculation:

[0109] First, sort the predicted values ​​from largest to smallest to get: S 12 =0.80, S 22 =0.75, S 32 =0.60.

[0110] The fuzzy measure parameters obtained from the system learning are: g({2}) = 0.40, g({2,3}) = 0.75, g({1,2,3}) = 1.

[0111] Substituting into Choquet's integral formula:

[0112]

[0113] Therefore, after fusing the three models, the predicted probability of this sample belonging to category 2 is 0.7325. Similarly, the same rule can be applied to perform fusion calculations for categories 1 and 3, ultimately yielding the fused output vector: The system will be Make a final classification judgment on the sample.

[0114] Through the aforementioned classifier construction and fuzzy fusion mechanism, this system significantly improves its ability to identify complex signal patterns while maintaining high accuracy, especially when the sample class distribution is uneven or the model predictions are highly divergent, demonstrating stronger stability and interpretability.

[0115] S5. After completing the fusion mechanism design, the system enters the joint training and optimization phase of the model. In this step, the system unifies all the aforementioned sub-modules, including statistical feature extraction, deep encoding, cross-attention fusion, three-class classifiers, and fuzzy metric parameters, into an end-to-end differentiable network structure to achieve joint optimization.

[0116] To ensure stable training and convergence of the model, the system uses the cross-entropy loss function as the primary optimization objective. For optimization, the system employs the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -5 To control overfitting, the model updates parameters in mini-batches of 64 during training, with a maximum of 100 training epochs. If the performance does not improve on the validation set for 10 consecutive epochs, training is stopped early (early stopping strategy).

[0117] After training, the system was evaluated on a reserved test set, with metrics including: classification accuracy, precision, recall, F1 score, and AUROC (area under the curve) for each class. Table 1 shows the ECG classification results on the PTBXL dataset. The proposed model outperformed the comparison method in all performance metrics.

[0118] Table 2. ECG classification results on the PTBXL dataset.

[0119]

[0120] The above results demonstrate that the present invention can achieve stable and efficient end-to-end learning and inference performance in multi-lead electrocardiogram signal classification tasks, and has significant engineering practical value.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An electrocardiogram (ECG) classification method based on multimodal feature fusion and fuzzy integral decision-making, characterized in that, Includes the following steps: S1. Acquire and preprocess the raw multi-lead electrocardiogram (ECG) signals, and extract time-domain and frequency-domain statistical features through discrete wavelet transform; S2. Use neural networks to perform end-to-end modeling of ECG signals and extract deep features across leads and across time. S21. The preprocessed signal in S1 For the signal of each lead Divide the data into N patches, each patch having a length of p. S22. Project each patch onto a fixed dimension using a fully connected layer. This forms an embedded feature representation; S23. Design two embedding mechanisms to enhance input structure information: Position embedding To distinguish the temporal location of a patch in the lead signal, a learnable vector and lead embedding are used. Assign a unique, learnable vector to each lead to represent its spatial location; S24. Concatenate all lead patches into an input tensor. The input is fed into the Transformer encoder module to construct a multi-layer encoder for modeling intra-patch temporal dependencies and inter-patch cross-lead spatial correlations, where... The dimensions to be embedded after splicing; S25. After encoding, the output vectors of each lead are aggregated using average pooling to form the representation vector of each sample. As the output of the depth modeling module, S3. To represent the dimension of the vector; Statistical features and deep features are fused through a bidirectional cross-attention mechanism to obtain a joint feature representation; S4. Construct multiple classifiers based on statistical features, deep features, and fusion features respectively, and fuse the multiple classification results using Choquet fuzzy integrals. S5. Based on the trained and optimized end-to-end model, automatically classify and diagnose cardiovascular diseases.

2. The electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision-making according to claim 1, wherein, In step S1, the statistical feature extraction includes the following steps: S11. Perform multi-level discrete wavelet decomposition on the ECG signal of the second lead to obtain several sets of detail coefficients and approximation coefficients. S12. Extract a set of clinically interpretable features from the original signal and the wavelet reconstructed signal, including RR interval, PR interval, QRS duration, QT interval, heart rate, heart rate variability, power spectral entropy, Shannon entropy, sample entropy, complexity, kurtosis, and skewness. S13. Stack the features from step S12 in a preset order to form a feature matrix. And it is processed using Z-score normalization, where It is a dimension of statistical characteristics.

3. The electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision-making according to claim 1, wherein, In step S2, the extraction of deep features includes: dividing each lead signal into multiple segments of length p, and mapping each segment to a fixed dimension through a fully connected network. The system adds learnable position embedding vectors and lead embedding vectors to each segment, combines them into an input tensor, and inputs it to a multi-layer Transformer encoder for modeling. The deep features obtained by average pooling the encoder output are represented as follows: As the output of the depth modeling module, represents the dimension of the vector.

4. The electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision-making according to claim 1, wherein, Step S3 achieves multimodal fusion through a bidirectional cross-attention mechanism, including the following steps: First, statistical characteristics with depth features They are mapped to query, key, and value respectively through linear transformations, specifically as follows: ； ； in These are queries obtained through linear mapping of statistical features, representing key and value pairs. A learnable matrix that represents a linear mapping of statistical features; These are the query results obtained from deep features through linear mapping, representing key and value pairs. It is a learnable matrix that represents a linear mapping of deep features; Then, the attention representation of statistical features on deep features is calculated. : ； And the attention representation of statistical features by deep features : ； in The feature representation dimension is obtained by linear transformation of statistical features and deep features. It is a normalized exponential function; Finally and After function After concatenation, a learnable weight matrix is ​​used. Projection as a fusion feature ,Right now: 。 5. The electrocardiogram classification method based on multimodal feature fusion and fuzzy integral decision-making according to claim 1, wherein, In step S4, the fuzzy integral fusion uses Choquet integral to integrate the outputs of multiple classifiers, including: denoting the three prediction vectors obtained based on statistical features, deep features, and fusion features as follows: , Concatenate rows into a matrix C represents the number of disease categories; for the matrix Each column vector Generate a matrix by arranging elements in descending order of size. And construct a fuzzy measure function that satisfies boundary conditions and monotonicity. Finally, the fusion output for category j for: ; in , and Represents the weight matrix. This represents the model index corresponding to the i-th sorting position determined by matrix S. and Represent matrices respectively and The elements in.

6. A system for implementing the method according to any one of claims 1 to 5, characterized in that, include: A clinical knowledge-guided statistical feature extraction module; The Transformer encoder's deep feature extraction module; the bidirectional cross-attention multimodal fusion module; Classifier and fuzzy integral decision module; The clinical knowledge-guided statistical feature extraction module obtains time-domain and frequency-domain statistical features through discrete wavelet transform; The deep feature extraction module of the Transformer encoder obtains deep features that characterize signal patterns and structures. A bidirectional cross-attention multimodal fusion module fuses statistical features and deep features to obtain a joint feature representation; The classifier and fuzzy integral decision module combine the prediction results of statistical features, deep features, and joint features obtained by the classifier. The fuzzy integral decision module integrates the multiple classification results to obtain the final prediction result.

Images