Research And Implementation Of End-to-end Ecg Classification Based On Deep Learning

Because of its simple and non-invasive nature,electrocardiography has become the most widely used technique for monitoring heart conditions.Subsequently,a computer-based assisted ECG diagnostic technique,namely the ECG classification algorithm,has been developed.In order to reduce the diagnostic workload of medical experts,ECG classification algorithm is a very valuable research direction.Traditional ECG classification methods require a lot of pre-processing work,such as denoising and removing baseline drift.The steps are cumbersome and the efficiency of the algorithm is low.At the same time,the increase of steps also increases the uncertainty in the operation of the algorithm.Therefore,this paper designs an end-to-end deep learning model to study the classification algorithm that directly processes the original ECG data without preprocessing,which increases the robustness of the algorithm while eliminating the algorithm steps.This paper first selects the MIT-BIH database to perform end-to-end classification of single-lead ECG data,and uses a smaller deep learning model to process single-lead ECG data.Aiming at the characteristics of electrocardiograms as timing signals,a PyTorch-based D-LSTM model was designed and compared with the traditional MLP and Vanilla-RNN models.The processing does not require pre-processing.Using the original ECG data,the designed D-LSTM model achieved a classification accuracy of 97.37%,which is 2% higher than that of the literature using the same database.Furthermore,we have studied the 12-lead ECG data classification algorithm.Because the multi-channel data is in a parallel synchronization relationship,it cannot be directly processed using the RNN structure.This article attempts a variety of data fusion methods based on different coding ideas.Four kinds of deep learning models based on PyTorch are designed: Multi-RNN,Vanilla-CNN,Feature-CNN andChannel-CNN.By comparing the classification results to analyze the characteristics of the ECG signal,the most suitable model for ECG signal classification is found out.The experimental CCDD clinical database was used.Compared with the single lead classification results of this database,the classification accuracy of the four multi-lead fusion models has been significantly improved.The classification accuracy of Multi-RNN is better than that of other RNN models.The substantial increase has proved the feasibility of the RNN structure in the multi-lead ECG classification algorithm.The optimal Channel-CNN end-to-end coding fusion model achieves the highest classification accuracy of 83.9%.