Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning

Researchers from arXiv introduce xMAE, a masked autoencoder pretraining framework that exploits the causal temporal relationship between ECG and PPG signals, where PPG lags ECG by vascular transit time, as a structured self-supervised constraint. Unlike prior SSL methods treating multi-modal biosignals as interchangeable views, xMAE enforces directional cross-modal reconstruction, yielding representations that outperform both unimodal and standard multimodal baselines. Wearable health AI developers and clinical ML engineers building on ECG/PPG foundation models should prioritize this approach for downstream cardiac monitoring tasks.