Stress Testing a Physics-Constrained Arterial-Pressure Twin Candidate: Reconstruction, Forecasting, Uncertainty, and Identifiability from Photoplethysmography

This is the authors' abstract. We don't add key points for this paper.

Calibration-free reconstruction of the arterial blood-pressure (ABP) waveform from photoplethysmography (PPG) is limited by a non-unique inverse mapping, source shift, timing uncertainty, and weakly identifiable physiology. Rather than proposing a new estimator, this study presents a systematic stress-testing framework and applies it to a physics-constrained arterial-pressure twin candidate assembled from established components (convolutional encoder-decoder, diastolic Windkessel constraint, observation-only recurrent state update, deep ensemble, and conformal calibration), using 61,795 synchronized PPG, ECG, and invasive-ABP segments under subject-disjoint, calibration-free evaluation. In the prespecified locked-test comparison, the Windkessel-constrained model showed no advantage over its capacity-matched baseline: the paired full-waveform MAE difference was + 0.091 mmHg (95% CI, − 0.071 to + 0.240 ), excluding the prespecified 5% improvement target, and the effect changed sign across validation seeds. An exploratory physics-weight sweep reduced the raw physics loss by 78.3%, confirming an active constraint, yet yielded at best a 0.78% validation gain; neither predictive benefit nor evidence of physiological validity. The causal forecaster outperformed last-beat persistence but trailed simple reconstructed-beat baselines by 0.89 - 1.36 mmHg at every horizon. Conformal bands reached 90.03% coverage only at a mean width of 242.01 mmHg, with weak error-uncertainty association. Temporal warping doubled MAE with almost no uncertainty response, high systolic pressures were strongly attenuated, cross-source transfer increased error by 61 - 86%, and only 5.6 - 7.7% of diastolic fits were practically identifiable. The candidate fails as a functional arterial-pressure twin; the contribution is a reproducible protocol showing where predictive, uncertainty, and mechanistic claims fail under independent tests.