CardioJEPA: A Framework for Holistic Cardiac Representations Through Predictive Multimodal Alignment with ECG-Preservation

Abstract The electrocardiogram (ECG) is the most widely used cardiac diagnostic tool, but it reflects cardiac structure and function only indirectly. Aligning ECG and cardiac magnetic resonance (CMR) representations offers a way to transfer imaging-derived structural knowledge into the ECG encoder without compromising the electrophysiological information it already carries. Contrastive learning has become the predominant approach for ECG--CMR alignment, yet the quality of the learned representation is jointly determined by the CMR target representation, the alignment objective, and the preservation of electrophysiological information unique to the ECG. We propose CardioJEPA, a predictive alignment framework built on frozen cardiac foundation models that addresses all three: a target that separates static anatomy from cardiac motion, a predictive objective that maps ECG representations onto a fixed CMR target without instance discrimination, and a preservation term that retains nearly all pre-alignment electrophysiological information. On 63,193 paired ECG--CMR recordings from the UK Biobank, CardioJEPA outperforms previous ECG--CMR alignment methods across structural, electrical, and diagnostic tasks, retains 98.6% of the ECG-unique information present before alignment, substantially reduces the modality gap relative to matched contrastive alignment, and transfers to external cohorts for structural heart disease and Chagas cardiomyopathy without task-specific training, achieving state-of-the-art performance.