Recognition Of ECG Rhythm And Morphological Abnormality By Combining Deep Learning And Decision-making Mechanism

The death rate of cardiovascular diseases in China ranks first among all causes,and arrhythmia is an important group of cardiovascular diseases.The standard 12-lead ECG signal is an important tool for diagnosing arrhythmia.Automatic detection of arrhythmias based on ECG signals are of great significance for the prevention and treatment of cardiovascular diseases.Although 12-lead ECG signal provides more comprehensive arrhythmia information than single-lead ECG.It is difficult to effectively merge the information of different leads,so it is still a challenging task to automatically detect arrhythmia with high accuracy and strong generalization ability with 12-lead ECG.In a 12-lead ECG signal containing arrhythmia,not all heartbeat beats are abnormal.Features related to arrhythmia only exist in heartbeat beats containing arrhythmia.12-lead ECG signals simultaneously record cardiac electrical activity from multiple spatial angles and there is correlation between different leads.Based on the above characteristics of the12-lead ECG signal,we first proposed the use of Squeeze-and-Excitation(SE)module for12-lead ECG signal arrhythmia detection.An end-to-end 12-lead ECG classification algorithm based on SE module and one-dimensional convolutional residual network is proposed.The SE module in the algorithm explicitly models the channel feature response to adaptively calibrate the interdependence between channels,enhance discriminative features and suppress noise,which can adaptively fuse feature information between different leads.The one-dimensional convolution operation along the time dimension is used to extract temporal features in the ECG signal,and the residual module makes the model easier to optimize.Because the convolutional neural network cannot process variable-length signals,in this paper,all ECG signals are padded zero to equal-length signals,and then used as the input of the model,which may cause interference to the model.To this end,an algorithm based on SE module and U-Net is proposed.According to the encoding-decoding idea in U-Net,the input ECG signal is first encoded into a feature space,and then decoded to obtain features for classification.The 12-lead ECG data set of the 2018 China Physiological Signal Challenge(CPSC2018)is used to verify the performance of the model.This data set is a real 12-lead ECG signal from 11 hospitals.The average F1 scores of the two models proposed in this paper on the test set are 0.828 and 0.824 respectively,which exceeds the performance of the existing algorithms.Experimental results show that our algorithm has good performance and has great practical application potential.