Synergy-Aware Contrastive Pretraining for Co-recorded Physiological Signals

Co-recorded electrocardiography (ECG) and photoplethysmography (PPG) waveforms encode complementary electrical and volumetric perspectives of cardiovascular function. Prevailing contrastive pretraining frameworks adopt pairwise or leave-one-out alignment between per-channel views, capturing shared information but failing to recover the synergistic interactions that emerge only when channels are processed jointly. We present SyneCo, a synergy-aware contrastive pretraining framework that integrates per-channel embeddings through an attention-based fusion block before applying the contrastive objective on the fused representation. A dual InfoNCE loss supervises the fused representation and per-channel components jointly, and channel dropout confers robustness against incomplete recordings. Pretrained on 3.92 million co-recorded segments from MIMIC-III, MIMIC-IV, and VitalDB, SyneCo consistently surpasses pairwise and leave-one-out baselines on six ECG, four PPG, and three multi-channel benchmarks, with the largest gains on patient attribute prediction tasks that depending on cross-channel synergistic correlates.