Physiologically Informed Attention: Tackling Class Imbalance in Myocardial Infarction Localization with an ST-Segment Guided Transformer

Accurate localization of Myocardial Infarction (MI) from 12-lead electrocardiograms (ECGs) faces two key challenges: the subtlety of ischemic patterns, which makes them difficult to detect, and the severe class imbalance inherent in clinical data. To address these issues, we propose ST-Former, a novel dual-path deep learning framework. At its core is a physiologically informed ST-Attention mechanism that dynamically identifies and focuses on critical pathological ST-segments. This specialized ST-guided path operates in parallel with a global Transformer path that captures the overall signal context. The two representations are then fused with clinical features (age, heart rate) for the final classification. On the PTB-XL dataset, ST-Former demonstrates robust performance in multi-label MI localization, with ablation studies confirming that its physiologically guided attention is a fundamentally more effective strategy for combating class imbalance than conventional methods.