Pulse2View: Same-Sensor Cross-View Pre-Training for Patient-Level Multi-Disease Classification from Wearable PPG

Anonymous (ICLR 2027 double-blind)

Summary

Aligns two views of each 30 s PPG recording with symmetric InfoNCE: a Ricker-CWT three-channel image (PPG + 1st/2nd differences) per 10 s segment, and a Conv1D-BiGRU temporal view; both encoders fused downstream, inference PPG-only. Pretrained on 99,701 unlabeled 30 s windows; evaluated on the private hospital-derived Clinical Wearable PPG Cohort (CWPC; 14,811 patients, 100,657 windows): highest three-seed mean on all four primary patient-level metrics — CHD AUROC 0.7859 ± 0.0044, Macro AUROC 0.7150 ± 0.0012 (+3.2 / +1.4 pts over ImageNet-style ConvNeXtV2-Tiny init; +2.6 / +0.8 over MOMENT-small). Gains hold at 25% / 50% label budgets.

Reported benchmarks