The Analysis Of ECG Signal Based On Deep Learning

Electrocardiogram is a weak electrical signal of cardiac activity recorded by an electrocardiograph,which can be used to diagnose arrhythmias such as atrial fibrillation and premature ventricular contractions in patients,in order to prevent them from causing sudden cardiac death.Human interpretation of ECG data is not only cumbersome and large,but also subject to limitation.With the continuous development of artificial intelligence technology,ECG signal analysis based on deep learning has attracted more and more attention,and has gradually become an effective means of clinical auxiliary diagnosis and treatment.Based on deep learning theory,this thesis took the MIT-BIH arrhythmia database as the research object,and conducted a classification study on three types of arrhythmia signals of supraventricular ectopic heartbeat,ventricular ectopic heartbeat,and fusion heartbeat and normal signals.The specific research contents are as follows:(1)In the preprocessing stage,the noise pollution and data categories imbalance in the MIT-BIH database were deeply studied.Firstly,according to the difference in frequency between ECG signal and noise,median filtering and wavelet transform were used to effectively filter low and high frequency noise in the signal;secondly,in the MIT-BIH database,the data were not evenly distributed,and the difference between the number of positive and abnormal samples was large,which also causes the phenomenon of false skew.The SMOTE oversampling method was used to oversample the three types of samples of supraventricular ectopic heartbeat,ventricular ectopic heartbeat and fusion heartbeat to alleviate the problem of overfitting caused by data imbalance.(2)In terms of network construction,the MISEResNet(Multiple scale Improve Sequeeze-and-Excitation ResNet)network in this thesis was constructed based on the residual network.Firstly,in order to avoid that the single-scale neural network cannot extract enough rich ECG signal feature information,a multi-scale residual network model MResNet was designed.This network model diverse scale features can be extracted;Secondly,considering the different importance of each scale of the ECG signal and the difference between the residual output features of each layer and the global information,this thesis draw on the idea of network compression the channel attention mechanism was introduced,and BN and Dropout were added to the traditional SE module to optimize its performance;the improved ISE module was applied to the multi-scale layer and the residual module respectively to extract more representative features.Finally,the residual module of traditional ResNet was adapted to place the second BN layer of the basic residual block in front of the convolution layer,further achieving uniformity in the input data distribution.(3)Considering the contextual features of ECG signals as time series signals,the MISEResNet-BiLSTM model was formed by combining the bidirectional long and short-term memory network with the MISEResNet in this thesis.Firstly,the MISEResNet model was used for ECG waveform feature extraction and data compression to establish the connection between waveforms;Then,the superimposed bidirectional long and short-term memory network was used to perform sequence analysis and time information extraction on the deep-level ECG signal feature sequence;Finally,based on MIT-BIH database has conduct a large number of experimental comparisons,and the results show that the network model was superior to some existing network models in waveform feature and time series feature extraction,with an accuracy rate of 99.31%.