Research And System Implementation Of ECG Classification Method Based On Deep Learning

Deep learning has been developed so far,academics have gradually shifted their attention to the study of ECG signals by combining traditional methods with deep learning,because the extensive digitization of ECG signals has promoted the application of deep learning on ECG signals.The traditional standardized interpretation of ECG signals contains three parts: signal preprocessing,manual feature extraction,and ECG signal classification,each of which is independent and interdependent,for example,the generalization of manually designed features is closely related to the performance of ECG signal classification models.Deep learning-based ECG signal classification can be trained end-to-end without relying on signal preprocessing and manual feature selection.However,the classification accuracy of these existing deep learning algorithms decreases significantly in practical applications,which makes it difficult to be applied in clinical experiments,and the final diagnosis still needs to be done manually by doctors.Based on the above status quo,the abnormal ECG classification method implemented in this thesis based on deep learning consists of the following three main parts:(1)Since the existing ECG classification network has poor feature extraction ability and cannot cope with multilead ECG signals well,this thesis proposes a multimodal ECG classification network with an end-to-end deep learning method to predict various dif-ferent arrhythmia types in multilead ECG by learning the common features among ECG multilead data and effectively using factors such as patient age and gender to achieve the improvement of the accuracy of ECG signals.(2)The multimodal ECG-based classification network itself does not explicitly learn the correlation between different leads and lacks the mining of information on the scale of ECG signals.This thesis proposes to use a multiscale network based on the attention mechanism to explicitly learn the relationship of multilead ECG signals,and to better learn the connection between ECG signal features by using parallel networks with different convolutional kernels to combine different scale features at the same spatial location.(3)The application of the algorithmic model to clinical diagnosis requires the sup-port of engineering direction.In this thesis,we design and implement an ECG abnormal-ity classification system,which transmits the ECG signals collected from PC to the cloud server via LAN,and the ECG abnormality classification model in the cloud server pro-cesses the ECG signals,and finally completes the functions of ECG drawing and display,ECG signal classification,classification result display,and case report writing.In this thesis,adequate experiments are conducted on publicly available ECG datasets such as MI-TBIH,and the multiscale network based on attention mechanism shows su-perior performance,surpassing other deep learning methods in several indicators such as F1＿Score,and the designed and implemented ECG abnormality classification system can initially meet the needs of clinical diagnosis.