Self-Supervised JEPA Pretraining for ECG Representations with Aligned Clinical Reports

Labeled electrocardiograms (ECG) are scarce, while unlabeled recordings are abundant. We pretrain a joint-embedding predictive architecture (JEPA) on unlabeled ECG segments by predicting the embeddings of masked leads and segments, and we align the resulting representations with clinical report text. We evaluate on a public 12-lead benchmark with a linear probe and a low-label regime, and on synthetic ECG with controlled anomalies. This is a research protocol, not a clinical tool. A synthetic pilot is too easy to discriminate, so the pretraining results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.