CN-122025106-A - Electrocardiogram multi-classification method and system based on deep learning

CN122025106ACN 122025106 ACN122025106 ACN 122025106ACN-122025106-A

Abstract

The invention relates to an electrocardiogram multi-classification method and system based on deep learning. The method comprises the steps of obtaining original electrocardiogram ECG data, preprocessing the data, adopting a data enhancement and oversampling method to conduct data processing to obtain ECG data samples, constructing a dual-channel feature extraction module composed of multiple-lead branches and single-lead branches, inputting and extracting global time sequence dependent features and local fine granularity features of the ECG data samples, fusing the global time sequence dependent features and the local fine granularity features to obtain fused electrocardiogram characterization, processing the electrocardiogram characterization based on a dynamic query generation mechanism to generate dynamic query corresponding to the original electrocardiogram ECG data, collecting noise data to construct denoising query, and conducting feature interaction and dynamic mapping in a transducer decoder to obtain an electrocardiogram multiclass prediction result. The adaptability to individual differences and noise can be improved, the robustness can be maintained under the noise and unbalance conditions, and the electrocardiograph classification is quick and accurate.

Inventors

WU YUANYUAN
SHI YANLI
HUANG MENGXING
FENG ZIKAI
ZHANG YU
ZHANG YUHANG

Assignees

海南大学

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (9)

1. An electrocardiographic multi-classification method based on deep learning, the method comprising: Acquiring original electrocardiogram ECG data for preprocessing, and adopting a data enhancement and oversampling method for data processing to obtain ECG data samples; Constructing a dual-channel feature extraction module consisting of a multi-lead branch and a single-lead branch, inputting the ECG data sample into the dual-channel feature extraction module, and performing convolution operation on the multi-lead branch to extract global time sequence dependence features of transconductance and performing attention mechanism on the single-lead branch to extract local fine granularity features; processing the electrocardiogram characterization based on a dynamic query generation mechanism, generating a dynamic query corresponding to the original electrocardiogram ECG data, and acquiring noise data to construct a denoising query; And inputting the dynamic query and the denoising query into a transducer decoder to perform characteristic interaction and dynamic mapping, so as to obtain an electrocardiogram multi-category prediction result.
2. The deep learning based ECG multi-classification method of claim 1, wherein the obtaining of raw ECG data for preprocessing and data processing using data enhancement and oversampling methods to obtain ECG data samples comprises: acquiring original Electrocardiogram (ECG) data, and carrying out normalization, length unification and filtering denoising processing on the original ECG data to obtain processed ECG data; And carrying out enhancement processing on the processed ECG data by adopting a MixUp data enhancement method, and increasing the number of samples in the processed ECG data by adopting a random oversampling technology to obtain ECG data samples.
3. The deep learning based electrocardiogram multi-classification method of claim 1 wherein inputting the ECG data samples into the dual channel feature extraction module and convolving the multi-lead branches to extract global timing-dependent features of the transconductors and the single-lead branches to extract local fine-grained features, comprises: Converting the ECG data samples into a matrix with uniform dimensions, and respectively inputting the matrix into a multi-lead branch and a single-lead branch; Capturing a time dimension global change mode of the matrix through time sequence convolution by a convolution network in the multi-lead branch, establishing association between leads by combining transconductance convolution and pooling, and outputting global time sequence dependency characteristics after activating functions and characteristic dimension reduction; Splitting the matrix into each lead signal according to leads through the single lead branch, inputting each lead signal into a feature extraction unit of an attention layer one by one, calculating time step association weight to highlight fine abnormal waveform information, and outputting local fine granularity features after nonlinear transformation, feature aggregation and splicing integration.
4. The deep learning based electrocardiogram multi-classification method according to claim 1, wherein processing the electrocardiogram characterization based on a dynamic query generation mechanism generates a dynamic query corresponding to the original electrocardiogram ECG data comprising: inputting the fused electrocardiogram characterization into a dynamic query generation module, and initializing a group of learnable query vectors of which the dimensions are matched with the electrocardiogram characterization and correspond to potential diagnosis categories; and performing attention interaction or nonlinear transformation on the electrocardiogram characterization and the leachable query vector based on a dynamic route correction mechanism to generate a dynamic query corresponding to the original electrocardiogram ECG data.
5. The deep learning based electrocardiogram multi-classification method of claim 4, wherein collecting noise data constructs a denoising query comprising: Collecting various environmental noises in a clinical scene, and sorting the environmental noises into noise data according to standard deviation and a duty ratio range; and randomly selecting part of queries from the dynamic queries, and superposing the noise data according to preset noise intensity and duty ratio to perform disturbance processing to construct and obtain denoising queries.
6. The deep learning based electrocardiogram multi-classification method according to claim 1, wherein inputting the dynamic query, denoising query into a transducer decoder for feature interaction and dynamic mapping comprises: carrying out dimension unified processing on the dynamic query and the denoising query, wherein the processed query is consistent with the characteristic dimension of the electrocardiogram characterization; Inputting the two types of processed queries and the electrocardiogram characterization into a multi-head attention layer in a transducer decoder, and calculating the association weight between the two types of queries and the electrocardiogram characterization; obtaining multidimensional depth interaction characteristics according to the association weights; performing layer normalization processing on the multidimensional depth interaction features, and generating intermediate feature representation after feature enhancement; And adaptively adjusting the mapping direction and intensity of the intermediate feature representation through a dynamic mapping mechanism of the transducer decoder to obtain an adjusted feature.
7. The deep learning based electrocardiogram multi-classification method according to claim 6, further comprising: Inputting the adjusted characteristics to a classification layer of the transducer decoder, and calculating the prediction probability of various diagnosis categories through a classifier; and acquiring a preset probability rule, judging category attribution according to the prediction probability based on the preset probability rule, and obtaining an electrocardiogram multi-category prediction result.
8. The deep learning-based electrocardiogram multi-classification method applied to an electrocardiogram multi-classification depth model according to claim 1, wherein the electrocardiogram multi-classification depth model comprises a dual-channel feature extraction module, a dynamic query generation mechanism module and a transducer decoder; the method further comprises the steps of: constructing a composite loss function to train the electrocardiogram multi-classification depth model, and realizing electrocardiogram multi-classification prediction through the trained electrocardiogram multi-classification depth model; the composite loss function comprises classification loss, cost sensitivity penalty term and denoising supervision loss, and is obtained by weighted summation of the classification loss, the cost sensitivity penalty term and the denoising supervision loss.
9. An electrocardiographic multi-classification system based on deep learning, the system comprising: The data processing module is used for acquiring the original electrocardiogram ECG data for preprocessing, and adopting a data enhancement and oversampling method for data processing to obtain ECG data samples; The feature extraction module is used for constructing a dual-channel feature extraction module consisting of a multi-lead branch and a single-lead branch, inputting the ECG data sample into the dual-channel feature extraction module, carrying out convolution operation on the multi-lead branch to extract global time sequence dependence features of transconductance and carrying out attention mechanism on the single-lead branch to extract local fine granularity features; the query generation module is used for processing the electrocardiogram characterization based on a dynamic query generation mechanism, generating a dynamic query corresponding to the original electrocardiogram ECG data, and acquiring noise data to construct a denoising query; And the category prediction module is used for inputting the dynamic query and the denoising query into a transducer decoder to perform feature interaction and dynamic mapping so as to obtain an electrocardiogram multi-category prediction result.

Description

Electrocardiogram multi-classification method and system based on deep learning Technical Field The invention relates to the technical field of artificial intelligence and medical health, in particular to an electrocardiogram multi-classification method and system based on deep learning. Background Cardiovascular disease remains a major cause of death and disability worldwide, and early diagnosis and risk prediction are critical to improving patient prognosis. An Electrocardiogram (ECG) is used as a noninvasive examination tool which is most widely applied in clinical practice and has high cost efficiency, can effectively detect abnormal heart electrical activity and evaluate heart function states, and is a physiological signal basis which is rich and key for screening and diagnosing heart abnormalities and diagnosing cardiovascular diseases. The traditional electrocardiogram interpretation is highly dependent on the clinical experience of a professional doctor, so that a great deal of time and effort are consumed to analyze and judge complex waveform signals, the traditional electrocardiogram interpretation has a certain subjectivity, and the high-efficiency diagnosis requirements under the conditions of large-scale clinical screening and basic medical resource shortage are difficult to meet. In recent years, with the development of deep learning, a great deal of research has explored the use of Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and transducer-based models in ECG automatic analysis. However, the existing ECG analysis method is mostly directly based on the integral input of the 12-lead ECG, so that diversity and complementarity among different leads are often ignored, and the global dependence of the transconductance and the fine granularity characteristic of a single lead are difficult to be simultaneously considered. In addition, ECG signals are often affected by noise interference and class imbalance in practical applications, resulting in reduced predictive performance of the model in small sample classes and abnormal modes, and lack of an effective modeling mechanism for individual differences and noise robustness, while dynamic, adaptive feature mapping is difficult to achieve in diagnostic tasks. Therefore, the traditional ECG analysis method depends on manual experience, and deep learning is difficult to achieve global and local characteristics, noise and category imbalance are prominent, and the problems of low ECG analysis accuracy and reliability often exist. Disclosure of Invention Based on the above, in order to solve the above technical problems, an electrocardiographic multi-classification method and system based on deep learning are provided, which can integrate transconductance and single lead characteristics, have dynamic query modeling capability, and can maintain robustness under noise and imbalance conditions. An electrocardiogram multi-classification method based on deep learning, the method comprising: Acquiring original electrocardiogram ECG data for preprocessing, and adopting a data enhancement and oversampling method for data processing to obtain ECG data samples; Constructing a dual-channel feature extraction module consisting of a multi-lead branch and a single-lead branch, inputting the ECG data sample into the dual-channel feature extraction module, and performing convolution operation on the multi-lead branch to extract global time sequence dependence features of transconductance and performing attention mechanism on the single-lead branch to extract local fine granularity features; processing the electrocardiogram characterization based on a dynamic query generation mechanism, generating a dynamic query corresponding to the original electrocardiogram ECG data, and acquiring noise data to construct a denoising query; And inputting the dynamic query and the denoising query into a transducer decoder to perform characteristic interaction and dynamic mapping, so as to obtain an electrocardiogram multi-category prediction result. In one embodiment, the method for obtaining the original electrocardiogram ECG data for preprocessing and adopting a data enhancement and oversampling method for data processing to obtain ECG data samples comprises the following steps: acquiring original Electrocardiogram (ECG) data, and carrying out normalization, length unification and filtering denoising processing on the original ECG data to obtain processed ECG data; And carrying out enhancement processing on the processed ECG data by adopting a MixUp data enhancement method, and increasing the number of samples in the processed ECG data by adopting a random oversampling technology to obtain ECG data samples. In one embodiment, inputting the ECG data samples into the dual-channel feature extraction module, and extracting global timing dependent features of transconductors by convolving the multi-lead branches, extracting local fine-grained features by an attentio