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

Research Objective

The authors address the challenge of detecting myocardial substrate abnormalities—such as myocardial scars and infarctions—using 12-lead electrocardiograms (ECG). These conditions are critical predictors of heart failure and sudden cardiac death, but detecting them via ECG is difficult due to subtle, lead-dependent signal variations, high-dimensional data, and significant class imbalance (where abnormal cases are much rarer than normal ones). The researchers aim to develop a deep learning model that is both highly accurate and interpretable for clinical use.

Approach: MSAIC-Net

The researchers propose MSAIC-Net (Multi-Scale Attention and Imbalance-Aware Contrastive Network). The model architecture features:

• Multi-Scale Feature Extraction: It uses parallel atrous (dilated) convolutional branches to capture both fine-grained local waveforms (like QRS complexes) and broader, long-range temporal patterns.
• Channel Attention: The model adaptively reweights feature channels and leads, allowing it to focus on the most clinically informative parts of the ECG signal.
• Focal Supervised Contrastive Learning: To combat class imbalance, the authors introduce a novel loss function. This strategy forces the model to learn "compact" representations for each class, pulling similar samples together in the feature space while pushing abnormal and normal samples further apart, with extra focus on "hard" (difficult-to-classify) examples.
• Post-Training Interpretability: Instead of relying solely on internal attention weights, the authors use permutation importance. This technique measures how much the model's performance drops when specific leads are perturbed, providing a direct, performance-based quantification of which ECG leads are most influential in a diagnosis.