Lead-grouped multi-stage learning for myocardial infarction localization

The electrocardiogram (ECG) is a ubiquitous medical diagnostic tool employed to localize myocardial infarction (MI) that is characterized by abnormal waveform patterns on the ECG. MI is a serious cardiovascular disease, and accurate, timely diagnosis is crucial for preventing severe outcomes. Current ECG analysis methods mainly rely on intra- and inter-lead feature extraction, but most models overlook the medical knowledge relevant to disease diagnosis. Moreover, existing models often fail to effectively utilize the global spatial relationships within multi-lead ECGs, limiting their ability to comprehensively understand and accurately localize the complex pathological mechanisms of MI. To address these issues, we propose a knowledge-driven overlapping lead grouping method. Based on clinical diagnostic knowledge, we group the 12 leads according to their relevance to MI localization while retaining the full set of 12 leads as a unified group. Additionally, we design a multi-stage learning network that first extracts basic features through initial convolutional layer and progressive convolutional block, followed by SE-enhanced multi-scale residual block and positional Transformer block to gradually learn deeper intra- and inter-lead features. Furthermore, we propose a branch-level weighted feature integration mechanism to effectively fuse the features extracted from each group. The proposed method was thoroughly evaluated on the publicly available multi-label PTB-XL dataset and achieved over 80% prediction accuracy for MI localization tasks. The results demonstrated significant improvements across several metrics compared to current state-of-the-art methods, confirming its exceptional performance.