Self-supervised representation learning reveals explainable physiological structure in high-dimensional magnetocardiography

Artificial intelligence (AI) has shown strong performance in cardiology, but most approaches rely on sensing modalities whose physical limitations constrain available information. Magnetocardiography (MCG) records the cardiac magnetic field with less tissue distortion than surface potentials and may preserve higher-dimensional spatiotemporal electrophysiological structure. Here, we investigated whether combining MCG with self-supervised learning enables physiologically meaningful cardiac representations. We developed MCG2Vec, a contrastive encoder trained directly on raw 64-channel MCG recordings. Using recordings from 1732 consecutive patients, learned embeddings were evaluated with task-specific probes for multivessel coronary artery disease, reduced left ventricular ejection fraction, and paroxysmal atrial fibrillation risk from sinus-rhythm recordings. The representations enabled discrimination of multivessel coronary artery disease (area under the receiver operating characteristic curve (AUC) 0.89), reduced left ventricular ejection fraction (AUC 0.81), and atrial fibrillation risk (AUC 0.77). Attribution analyses revealed probe-specific temporal and spatial patterns corresponding to ventricular depolarization, repolarization, atrial activation dynamics, and coronary territories, supporting physiological interpretability. These findings suggest that higher-fidelity sensing combined with self-supervised representation learning can yield structured and explainable embeddings from non-invasive cardiac magnetic field recordings. More broadly, the study highlights measurement physics as an important determinant of what medical AI systems can learn.