Machine learning-enhanced MCG for LVH detection: a multi-domain feature selection approach

Magnetocardiography (MCG) provides high-resolution spatiotemporal insights into cardiac electrophysiology but remains underutilized for left ventricular hypertrophy (LVH) diagnosis due to a lack of interpretable analytical tools. We propose a novel interpretable machine learning framework that systematically decodes MCG signals across four complementary domains: temporal waves, spatial waves, current source imaging, and dynamic characterization. To address class imbalance, we integrated Focal Loss into the XGBoost objective function. In a dataset of 481 subjects, our model achieved an AUC of 0.902 in cross-validation and 0.837 in independent validation, significantly outperforming conventional baselines. Notably, SHapley Additive exPlanations (SHAP) identified T-wave magnetic polarity as the most influential predictor, offering new perspectives on the electrophysiological remodeling of hypertrophied myocardium. This framework bridges the gap between raw sensor data and clinical decision-making, providing a robust tool for automated LVH detection.