Electrocardiography (ECG) is a cornerstone for cardiac diagnosis, yet most machine‑learning approaches lack interpretability, hindering clinical adoption. We introduce IECG, a weakly‑supervised ECG foundation model that jointly performs diagnostic classification and abnormality localization without any manual region annotations, addressing a long‑standing bottleneck in interpretable ECG learning.
The key component of IECG is a weakly‑supervised attention mechanism that derives pseudo‑labels for abnormality heatmaps from established clinical knowledge. These pseudo‑labels supervise the CLS‑token attention via a KL‑divergence loss, encouraging the model to focus on diagnostically relevant waveform segments.
During pre‑training, IECG is optimized with three objectives: clinically guided attention supervision, ECG‑text contrastive alignment, and report captioning. This multi‑task setup enables the model to learn rich multimodal representations while producing localized waveform evidence.
By explicitly aggregating features from clinically guided attention regions, IECG grounds its diagnostic predictions in interpretable waveform evidence without sacrificing classification performance. Extensive experiments show that IECG achieves state‑of‑the‑art abnormal region localization while maintaining strong classification results, advancing trustworthy foundation models for ECG analysis.
Review