Research On ECG Signal Classification Algorithm Based On Deep Learning

Cardiovascular disease has always been the biggest cause of human death,Electrocardiogram(ECG)examination is an effective means of diagnosing cardiovascular disease,but its diagnostic process is difficult,subjective,and prone to misjudgment.In recent years,the usage of Convolutional Neural Network(CNN)to solve the problem of ECG classification has become a hot topic.However,the weakness of CNN lies in the lack of ability to extract timing features and its "black box" characteristics,which makes the algorithm effective but difficult to be accepted by the medical diagnostic field.Based on the above background,this thesis conducts research on ECG signal classification algorithm,and the main research contents include:(1)Aiming at the problem that traditional denoising methods are easy to cause oscillation phenomena in the abrupt position of ECG,and considering the characteristics of ECG noise,the translational invariant characteristics of Stationary Wavelet Transform(SWT)and the independent signal separation ability of Independent Component Correlation Algorithm(ICA)are used.The SWT-ICA method is designed for ECG denoising.By comparing the denoising capabilities of different methods on ECG signals,the effectiveness of the SWT-ICA method is verified.(2)In view of the problem that traditional machine learning algorithms excessively rely on manual feature extraction,1D-VGG16,1D-ResNet and 1DResNext algorithms suitable for ECG classification are designed based on classical convolutional neural network.Experiments are carried out on MIT-BIH dataset and PTB dataset respectively to analyze the influence of different algorithms and parameters on ECG classification.The results show that the algorithm designed in this thesis has good effect and generalization performance on ECG classification task,and can be used to assist doctors in diagnosing abnormal ECG signals.(3)The ResNext-CSA-GRU classification model is established.To solve the problem that CNN does not distinguish the importance of features,Channel and Spatial Attention(CSA)mechanism is introduced to learn different features with emphasis.In view of CNN’s lack of ability to extract time series features of ECG,it adds a Gate Recurrent Unit(GRU)structure after the classification model.The average sensitivity,accuracy and F1-score of this model on MIT-BIH data set are 99.34%,99.60%and 99.46%,respectively.The average sensitivity,accuracy and F1-score of this model on PTB data set are 80.22%,80.78%and 80.54%,respectively.The ResNext-CSA-GRU model designed in this thesis has achieved a better level in similar studies.In view of the "black box" characteristics of deep learning algorithm,the sample level visual analysis of the designed classification algorithm is carried out,which intuitively demonstrates the basis for the model to make judgment.Combined with the analysis of ECG diagnosis theory,the results show that the judgment basis of the ResNext-CSA-GRU model established in this thesis is consistent with the ECG discrimination criteria.The visualization research carried out provides new ideas for the follow-up ECG algorithm research.