MaskedGAFNet, a self-supervised learning framework using GADF representation, outperformed supervised baselines under limited labels for myocardial infarction localization.
Does MaskedGAFNet improve the accuracy of myocardial infarction localization on 12-lead ECGs compared to supervised baselines and prior self-supervised methods?
A novel self-supervised learning framework, MaskedGAFNet, improves the automated localization of myocardial infarction on 12-lead ECGs, offering a promising tool for interpretable ECG analysis.
Myocardial infarction (MI) remains a major cause of mortality worldwide, requiring rapid and accurate localization of infarct regions. While 12-lead ECGs are the clinical standard for MI diagnosis, interpretation is often expertise-dependent and inconsistent. To address this, we propose a novel self-supervised learning framework based on Gramian Angular Difference Field (GADF) representation, which converts 1D ECG signals into 2D images that capture temporal and morphological patterns. This enhances robustness to noise and aligns with clinical visual interpretation. Based on this representation, we introduce MaskedGAFNet, a self-supervised learning framework using masked image modeling to learn from unlabeled GADF images. Experiments on infarction localization show that MaskedGAFNet outperforms supervised baselines under limited labels and significantly improves over prior self-supervised methods. Grad-CAM visualizations further confirm attention to clinically relevant ECG segments. These results highlight the promise of GADF and self-supervised learning in interpretable ECG analysis.
Wang et al. (Sun,) conducted a other in Myocardial infarction. MaskedGAFNet vs. Supervised baselines and prior self-supervised methods was evaluated on Infarction localization. MaskedGAFNet, a self-supervised learning framework using GADF representation, outperformed supervised baselines under limited labels for myocardial infarction localization.