Research On Classification Algorithm Of ECG Signal Based On Deep Learning

As the pace of life accelerates and the work pressure increases,the number of heart disease cases is increasing year by year,seriously threatening people's health and life.If the patient's arrhythmia such as atrial fibrillation and ventricular premature beats can be timely diagnosed by characteristics of the shape and appearance of the waveform displayed on the electrocardiogram,it can prevent more serious heart disease and sudden cardiac death.The current number of ECG doctors is insufficient to deal with massive ECG,and doctors need computer-aided diagnostic classification to improve diagnostic speed.However,due to the weakness and complexity of ECG signals,there are many kinds of interferences and individual differences in the signal.It is difficult to achieve efficient ECG signal classification algorithm.Aiming at the problems above,this thesis proposed the ECG signal classification algorithm ResNet-Bi-LSTM based on residual network(ResNet)stacking Bidirectional Long Short Term Memory(Bi-LSTM).The proposed algorithm made full use of the residual network to extract the waveform characteristics of ECG signals and enhance the extraction ability of context features of signal sequences by stacking Bidirectional Long Short Term Memory.This algorithm overcame the shortcomings of manually designing and extracting features in traditional methods.The main research work is divided into the following sections:1.This thesis used median filtering combined with high-pass filtering,Butterworth band-stop filter,and wavelet threshold method to filter out the noise such as baseline drift,power line interference and myoelectric interference in the original ECG signal.Finally,high quality ECG signals were obtained for subsequent feature extraction and neural network classification.2.In this thesis,wavelet transform was used to identify the R-peak of the filtered ECG signal.The heartbeat template was extracted and the correlation coefficient between all the heartbeat templates and the standard template were calculated.Then the threshold was set to eliminate the heartbeat that is seriously interfered and the ECG signal was reconstructed.Heart beat template features and traditional waveform features were separately extracted on the heart beat templates and the reconstructed signal.Finally,XGBoost model based on all extracted features was constructed and tested on the 2017 PhysioNet/CinC Challenge dataset and clinical ECG dataset.3.In this thesis,ResNet-Bi-LSTM algorithm was proposed,residual network was designed to automatically extract the ECG signal features layer by layer,and Bidirectional Long Short Term Memory was superimposed to enhance the ability of extracting the ECG signal timing features.This thesis increase the number of samples by panning and randomly exchanging segments.The residual network structure of Global Average Pooling layer connecting Fully Connected Layer at the end and stacking 15 residual blocks was determined by cross validation.Then the Bidirectional Long Short Term Memory was superimposed to analyze sequence of deep ECG signal characteristics and extract the time information.Finally,this model obtained a classification performance superior to the traditional classification method on the 2017 PhysioNet/CinC Challenge dataset and clinical ECG signal dataset.The ResNet-Bi-LSTM algorithm had an average F1 of 0.8852 and omission diagnostic rate of 0.368%on 92,425 clinical ECG signal samples,which was much better than the traditional ECG signal classification algorithm.In addition,the Random Forest model,XGBoost model,residual network model,ResNet-Bi-LSTM model are combined by Stacking ensemble to further improve the clinical application performance.The final model had an average F1 of 0.8914,and the omission diagnostic rate was reduced to 0.071%.