Automated Detection and Localization of Myocardial Infarction With Interpretability Analysis Based on Deep Learning

Electrocardiogram (ECG) is a noninvasive, simplest, and fastest way to diagnose myocardial infarction (MI). Although different methods have been leveraged based upon deep learning covered by the existing studies, the spatial-temporal relationship in the lead and between leads has not been deeply analyzed. To address the issue, a novel multilead branch with the residual network integrated with squeeze and excitation networks and bidirectional long short-term memory (LSTM) model named MLB-ResNet-SENet-BL was presented. First, spatial features were exploited by the morphological information representation network in the lead based on MLB-ResNet. Then, these feature mappings among these spatial features based on SENet were strengthened and weakened by the importance analysis network of feature mapping in the lead, respectively. In addition, the temporal features were extracted by the dependency network between the leads based on BLSTM. Meanwhile, the model was evaluated using fivefold cross validation for MI detection and localization based on PTB and PTB-XL. The resulting model outperforms the state-of-the-art studies for intra-patient and inter-patient paradigms. The interpretability analysis using class activation mapping with gradient (Grad-CAM) was also leveraged for visualization of the specific QRS waves and ST-T segments of 12-leads ECG, which demonstrated that the highlighted parts of heat maps were completely in line with the diagnostic basis and strategy of doctors. Deployment of such models can potentially help ensure the life safety of patients and strive for the best treatment opportunity.