Multi-class 12-lead ECG automatic diagnosis based on a novel subdomain adaptive deep network

Arrhythmia is a common type of cardiovascular disease, which has become the leading cause of global deaths. Recently, automatic 12-lead ECG diagnosis system based on numerous labelled data has attracted increasing attention. However, labelling 12-lead ECG recordings is a complex and time-consuming task for clinicians. And then, the existence of data distribution differences limits the direct cross domain use of the trained model. Enlighted by subdomain adaptation methods, this paper designs a novel subdomain adaptative deep network (SADN) for excavating diagnosis knowledge from source domain dataset. Firstly, convolutional layer, residual blocks and SE-Residual blocks are utilized for extracting meaningful deep features automatically. Additionally, feature classifier uses these deep features for obtaining the final diagnosis predictions. Further, designing a novel loss function with local maximum mean discrepancy is utilized for restricting data distribution discrepancy from different datasets. Finally, the Clinical Outcomes in Digital ECG and 1st China Physiological Signal Challenge datasets are utilized for evaluating the superiority of SADN, which presents that SADN enhances algorithm performance on the unlabelled target domain dataset. Further, compared the existing methods, the proposed network structure acquires better performance with a F1-macro of 89.43% and a F1-macro1 of 87.09%. Besides, among the 4 kinds of ECG abnormalities, the diagnostic effect of the SADN is better than that of cardiology residents. Thus, SADN has promising potential as an auxiliary diagnostic tool for the clinical environment.