Electrocardiogram (ECG) is a non-invasive, simplest and fastest way to diagnose myocardial infarction (MI). Although different methods have been leveraged based upon deep learning covered by existing studies, the spatial-temporal relationship in the lead and between leads has not been deeply analyzed. To address the issue, a novel multi-lead branch with the residual network integrated with squeeze and excitation networks and bidirectional long short-term memory model named MLB-ResNet-SENet-BL was presented. Firstly, 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. Additionally, the temporal features were extracted by the dependency network between the lead based on BLSTM. Meanwhile, the model was evaluated using 5-fold 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 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.

Myocardial infarction (MI) boasts the highest mortality rate, striking suddenly and often stealthily. Conducting an electrocardiogram (ECG) examination stands as the most straightforward and rapid method to diagnose MI, rendering research into MI intelligent diagnosis via ECG of paramount importance. We developed search and selection strategy to extract articles from Google Scholar and PubMed, analyzed the existing studies of MI intelligent diagnosis algorithms using ECGs according to three typical aspects: MI data enhancement, MI diagnosis models and interpretability of models. Firstly, techniques for enhancing MI data were summarized, then handcrafted features and models with automatic feature extraction were discussed. Meanwhile, the interpretability of MI diagnosis models was analyzed including Grad-CAM, SHAP, multi-head attention mechanism and knowledge graph. A comparative analysis encompassing the methodologies, datasets, the number of leads and input modalities, performance evaluation, as well as interpretability for MI diagnosis was given. We found that deep learning has achieved good performance and knowledge graph accomplished the best interpretability. Finally, limitations and challenges were pointed out, including lack of high-quality MI records, absence of dynamic ECG records, effective representations of different ECG morphologies, the promotion of models under inter-patient paradigm, the improvement of transparency and interpretability of MI diagnosis, the high complexity of models. To our knowledge, this is the first comprehensive and multidimensional review analyzing intelligent diagnosis of MI using ECG signals.