CardioState-JEPA Learns Shared Cardiac Representation Across Modalities.

Existing cardiac foundation models typically focus on a single sensing modality, failing to exploit the shared physiological information present across different cardiac signals like Electrocardiography (ECG), Photoplethysmography (PPG), and Phonocardiography (PCG). This research presents CardioState-JEPA, a novel cardiac foundation model designed to learn a unified, shared representation from these heterogeneous signals. It employs a physiology-aware joint-embedding predictive architecture. CardioState-JEPA maps diverse waveforms into a common token space and processes them with a shared Transformer encoder. It learns by predicting masked latent cardiac states, focusing on shared physiology rather than sensor-specific appearances. To address the inherent temporal delays between electrical, mechanical, and hemodynamic events, the model incorporates a learned delay aligner. The training strategy first learns within-modality structure from abundant unimodal data, then aligns modalities using scarcer paired data. Evaluation across 25 downstream tasks shows significant improvements in classification and detection for PPG, PCG, and ECG compared to self-supervised baselines, matching or exceeding models trained with clinical text or supervised labels.