ECG Classification And Identification Based On Deep Learning

Traditionally,doctors used manual methods to interpret ECG to assess patients’condition.In recent years,artificial intelligence technology has promoted information mining of ECG data,achieved ECG automatic classification and abnormal findings,and realized ECG individual identification system as well.A typical AI-based ECG analysis system consists of three main processes,ECG sample generation,feature extraction,and classification.Existing research has obvious shortcomings in these three processes.Above all,sample generation,the fixed sequence length is used to segment the beat and the starting position of the segmentation is fixed,ignoring the dynamics of the beat length and resulting in the insufficiency of samples,which seriously affects the classification performance.Secondly,the different methods proposed to extract features are mostly considered from single beat scale,ignoring the local features of a beat and the intrinsic association between beats.Finally,the simple neural network used for classification cannot obtain accurate weights of different feature vectors,resulting in that the model classification accuracy is limited.In summary,with the utilization of deep learning,there is still a significant space for improvement.To this end,this thesis concentrates on the study of ECG based on deep learning,and completes two tasks,i.e.ECG classification(task 1)and ECG identification(task 2).The core idea of this thesis is to improve the performance of the ECG system by proposing accurate beat segmentation and effective feature extraction methods,improving neural network classifier,which manifests significant value both in theory and application aspects.The main contributions of the thesis are as follows.(1)An adaptive dynamic beat segmentation method is proposed to accurately segment the beat and improve the classification accuracy corresponding to task 1.The method uses dynamic RR interval length for representation of single beat length,through which effective segmentation and dynamic representation are realized simultaneously.In addition,task 2 improves a sample extraction method,which solves the problem of samples deficiency,and avoids the unreliability of high evaluation accuracy caused by the overlap of training data and test data in existing research.(2)A multi-scale feature extraction method is proposed.Significant features of ECG captured from multiple dimensions obviously improve the classification accuracy for task 1.The main idea of the method is that features scaling on the local beat,the single heart beat and the multiple beats are extracted automatically and manually,respectively.For the former,it is achieved with the use of deep learning,while for the latter,ECG medical knowledge is involved.Therefore,the expression of the sample achieves variety and accuracy at the same time.(3)The channel attention method in computer vision is introduced into the neural network we used,and weights are automatically assigned to different feature vectors,which further enhances the model accuracy.The final experiments show that the accuracy of this thesis in task 1 and 2 reaches 96.94%and 95.86%,which is 3.5%and 3.0%higher than related work,respectively.