MSAIC-Net: A Multi-Scale Attention and Imbalance-Aware Contrastive Network for ECG-Based Myocardial Substrate Abnormality Detection

Researchers present MSAIC-Net, a deep learning framework that improves ECG-based detection of myocardial substrate abnormalities like scarring and heart attacks. The model combines multi-scale attention mechanisms with contrastive learning to address class imbalance and interpretability challenges, demonstrating strong performance on both institutional and public datasets.

MSAIC-Net represents a meaningful advancement in clinical AI applications by tackling the specific challenge of detecting heart disease markers from electrocardiogram data. The framework addresses a critical healthcare problem: ECG remains one of the most accessible diagnostic tools globally, yet its diagnostic accuracy for subtle abnormalities remains limited. Traditional deep learning approaches struggle with the heterogeneous nature of multi-lead ECG signals and the natural class imbalance in medical datasets where normal cases far outnumber abnormal ones.

The technical innovation lies in combining three complementary approaches. Multi-scale atrous convolutions capture temporal patterns across different time horizons, while channel attention mechanisms identify which ECG leads and features matter most for detection. The imbalance-aware contrastive learning strategy directly addresses a persistent problem in medical AI: models trained on imbalanced data often perform poorly on the minority abnormal class that clinicians most care about detecting.

The inclusion of lead-wise permutation importance analysis addresses a critical gap in clinical AI adoption. Healthcare practitioners require interpretability to validate whether models rely on clinically meaningful signals rather than spurious correlations. By quantifying each ECG lead's contribution, the framework builds trust and enables clinical validation.

The dual evaluation across a small institutional cohort and a large public dataset demonstrates practical value. Small hospitals can benefit from transfer learning even without extensive labeled data, while larger institutions can leverage the PTB-XL benchmark validation. This accessibility matters for healthcare AI adoption, where many facilities lack data science infrastructure.

• →MSAIC-Net combines multi-scale convolutions and contrastive learning to improve ECG-based detection of myocardial abnormalities with strong interpretability.
• →The framework specifically addresses class imbalance in medical datasets, a persistent challenge that limits minority class detection accuracy.
• →Lead-wise permutation importance provides clinicians with interpretable feature attribution, essential for regulatory approval and clinical adoption.
• →Performance gains are particularly pronounced in low-data scenarios, expanding clinical AI applicability to smaller healthcare institutions.
• →The dual evaluation on institutional and public datasets demonstrates both research rigor and practical deployment readiness.