Research Of Automatic Classification Method For Multi-lead ECG Based On Convolution Neural Network

With the rapidly development of human society and economy,the rhythm of people’s life is accelerating,causing people’s mental stress to increase sharply,and the induction of heart disease has become a common disease that poses a major threat to human health.Electrocardiogram is one of the basic routine diagnosis and treatment techniques for heart diseases in modern hospitals,which can provide important reference for clinicians.However,the research on the automatic classification of traditional electrocardiogram focuses on the classification of heart beat type.Although the accuracy of the test set is very high,the actual clinical effect is very poor.The final diagnosis still needs to be done manually by the doctor,indicating that the heart beat classification is not very feasible.Misdiagnosed and misjudgment occurs,which affects the treatment of heart disease.In recent years,the deep learning method represented by convolutional neural network has achieved remarkable results in the field of image classification processing.For this reason,this paper draws on its advantages in automatic feature extraction and classification accuracy,and focuses on multi-lead based on convolutional neural network.The electrocardiogram automatic classification method has carried out research work,and the research content mainly includes:(1)Research on preprocessing method of multi-lead ECG data.The ICBEB ECG database was used to contain 6877 ECG data,each of which was derived from 12-lead ECG data from different clinical patients.In order to improve the classification and recognition effect,for the types of noise such as baseline drift,myoelectric interference and power frequency interference existing in ECG data acquisition,corresponding filters are designed to filter out.(2)Research on Automatic Classification of Multi-lead ECG Based on Resnet50 Network.Different from the traditional ECG automatic classification,the design feature extraction algorithm is different from the classification process.In the research,the Resnet50 network in the convolutional neural network is used to perform deep neural network learning on the multi-lead ECG data,and the ECG features are automatically extracted.It is fused with the classification process,and finally the decision function is fitted to automatically classify the electrocardiogram,which includes normal heart rhythm(Normal),atrial fibrillation(AF),and primary atrioventricular block(I-AVB).),left bundle branch block(LBBB),ventricular early-onset contraction(PVC),right bundle branch block(RBBB),ST-segment depression(STD),atrial premature contraction(PAC),ST-segment elevation High(STE)and other clinical arrhythmia classification,obtained the average F1 = 0.7,and some abnormal indicators F1 = 0.86.(3)Resnet network optimization based on expanded convolution and automatic classification of electrocardiogram.Based on the Resnet residual network structure,the characteristics of the high-dimensionality of 12-lead ECG data are studied.By improving the expansion convolution module to increase the receptive field,reduce the number of network layers,optimize the loss function,and improve the network training steps.Optimize measures and test on the ICBEB ECG dataset.The experimental results show that the classification accuracy of the optimized model can reach 80%,and the evaluation index of 9 kinds of heart rhythm categories is F1=0.74,which is obviously improved compared with the unoptimized network performance.