MCG2Vec, a self-supervised learning model, detects multiple cardiac pathologies from magnetocardiography with interpretable patterns aligned to cardiac physiology.