Accurate but Not Transferable: Can we overcome Identity Confounding in Blood Pressure Regression? (MOrdereD)

Shows that PPG→BP regressors can post low pooled error on held-out subjects by encoding subject identity rather than BP change: subject-disjoint linear probes on published encoders recover at most R² = 0.11 of systolic BP change, losing 40–83% of that decodability on unseen subjects; error stratified by BP-change magnitude exposes severe failures above 60 mmHg that pooled error hides. MOrdereD constrains learning to within-subject change at three levels: (1) input pairs an event recording with a same-subject baseline and the target is the BP difference; (2) each batch holds one pair from each of B subjects; (3) a ranking objective over paired embeddings by BP change, plus a decoder reconstructing the event signal. Pretrained on finger PPG from ICU/surgical cohorts and transferred to wrist PPG in clinical and daily-life settings: linear decodability of BP change R² = 0.22 with only 9–14% loss on unseen subjects; clinical setting — 34% lower error than a zero-change predictor across change strata, 57% lower for changes >60 mmHg, pooled MAE 8.3 vs 11.0 mmHg, ≥9 pp better macro skill than all baselines; daily life gain 18%, large changes still poorly tracked.

• Form factor: wrist

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.