A Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models

Anonymous (ICLR 2027 double-blind submission)

Summary

Audits frozen representations of four encoders on 37,784 MC-MED encounters via concept decodability, covariate-adjusted and incremental prediction, and post-erasure predictability with encounter-matched ECG comparisons. AnyPPG decodes pulse-derived concepts well, but for potassium (6,348 matched encounters) its correlation drops 0.21 → 0.05 after age and renal adjustment, while ECGFounder retains 0.32 → 0.27 (paired adjusted diff +0.22, 95% CI [0.19, 0.26]). Creatinine also attenuates.

Reported benchmarks