Digital Twin Falls Short: Physics-Based Blood Pressure Model Fails Rigorous Stress Test

A rigorously designed stress test of a physics-constrained “arterial-pressure twin” candidate has delivered a set of carefully bounded negative results: the model implements every component that a cardiovascular digital twin is supposed to have, yet under locked, unseen-subject evaluation it earned none of the claims attached to the word twin. It does not.

The motivation is straightforward. Cuffless blood pressure estimation from photoplethysmography, the optical sensor ubiquitous in wearables and clinical monitors, has long promised continuous, unobtrusive monitoring. But recent evaluations have found the approach unreliable under ambulatory, calibration-free, and heterogeneous conditions, particularly for systolic pressure. Because PPG is not a pressure measurement, different pulse shapes can correspond to similar pressure levels, and standardized benchmarks show substantial performance losses when evaluation is subject-disjoint or transferred across acquisition sources. Meanwhile, the term “digital twin” has proliferated in healthcare faster than validation, often without persistent data connection, verification, or defined scope.

The researchers defined a functional arterial-pressure twin by five testable requirements: maintaining an individualized observation-updated state, reconstructing the current arterial blood pressure waveform, forecasting beyond the observation boundary better than trivial extrapolation, providing uncertainty that is both calibrated and sharp enough to inform decisions, and reporting mechanistic quantities only when practically identifiable. A model implementing the components without demonstrated satisfaction of these requirements is called a twin candidate. The candidate examined here implemented all five components, and failed requirements three through five.

The confirmatory comparison was a capacity-matched showdown. Two models with exactly 2,188,718 trainable parameters each shared an identical convolutional encoder-decoder backbone; the only difference was that one, P1, added an effective diastolic-state estimator and a Windkessel consistency loss, a physics constraint describing passive diastolic pressure decay. On a locked calibration-free test set of 3,781 segments from 279 unseen subjects drawn from the PulseDB benchmark of intensive-care and perioperative recordings, the physics-constrained model was slightly worse: a paired full-waveform mean absolute error difference of +0.091 mmHg, with a 95 percent confidence interval excluding the prespecified practical target of a 5 percent improvement. Across three paired training seeds on validation subjects, the effect averaged -0.012 mmHg and changed sign, showing the locked difference lies within the range of training stochasticity alone.

An exploratory sweep confirmed the physics term was genuinely active, reducing the raw physics loss by 78.3 percent at the strongest weight, yet the best validation improvement was below 1 percent and stronger weights degraded accuracy by 9.5 percent. Activity, the authors stress, shows optimizer response, not physiological validity. The identifiability audit made this concrete: only 5.61 to 7.74 percent of diastolic reference fits passed practical identifiability gates across settings, and within that subset the learned and reference time constants were negatively correlated, with errors as large as the quantities being estimated. Abstention, not mechanistic interpretation, was the honest real-data outcome.

Forecasting fared no better against strong comparators. The causal forecaster P2, which predicts complete waveforms one, three, and five beats ahead using only observations through the current beat, beat last-beat persistence by 17.6 to 21.1 percent. But persistence is a weak null. Against baselines built from the model’s own recent reconstructions, repetition of the previous cycle, the mean of recent beats, or repetition of the current reconstruction, P2 was worse in all twelve comparisons, with paired deficits ranging from +0.888 to +1.364 mmHg and confidence intervals excluding zero. Because the waveform is nearly stationary over one to five beats, “no change” is the appropriate null, and the model failed it. A capacity-matched ablation removing the recurrent state showed no added forecasting value.

Uncertainty quantification was calibrated but operationally useless. Segment-simultaneous conformal calibration achieved 90.03 percent nominal coverage only with a mean prediction band 242 mmHg wide, roughly four times the average pulse pressure in the cohort and wider than nearly the whole pressure range studied. Error-uncertainty rank correlations were weak, 0.129 at segment level and 0.254 at subject level, and rejecting the 20 percent most uncertain cases lowered error by only 1.8 percent. Under maximum temporal warping corruption, waveform error more than doubled, rising 102.4 percent, while predictive uncertainty increased by just 0.75 percent. The uncertainty layer, the paper concludes, supports failure analysis, not interval estimation or abstention.

Other stress tests exposed further fragility. Cross-source transfer between the MIMIC-derived and VitalDB-derived halves of PulseDB raised error by 61.0 and 85.6 percent relative to in-source performance, a gap attributable mostly to input-side shifts in sensor, timing, and clinical population rather than pressure-distribution shift. Alignment sensitivity was substantial, with a 100-millisecond lag perturbation raising error by roughly 5 to 6 mmHg. Adding electrocardiography as a second input channel increased error by about 1 mmHg. Most consequentially, systolic pressure was strongly attenuated: signed error fell by 0.796 mmHg for every 1-mmHg increase in reference systolic pressure, and systolic error climbed from 12.00 mmHg in the normal stratum to 63.49 mmHg in the extreme stratum, with 94 percent of hypertensive segments underestimated by more than 10 mmHg.

The authors are careful about scope. All data came from hospital recordings with invasive reference arterial lines and high signal quality, with no ambulatory or wearable data tested, so performance under motion, posture change, and variable contact pressure is expected to be worse. The negative results apply to this specific Windkessel formulation and information setting, not to all physics-informed cardiovascular models. No claims of device validation, diagnosis, or treatment decision support are made, and the candidate cannot be used clinically.

The defensible contribution is the framework itself: a reproducible template demanding that a mechanistic term be shown both active and beneficial against a capacity-matched model, that forecasters beat repetition of their own recent reconstructions rather than only persistence, that uncertainty be reported jointly as coverage, width, and error ranking, and that mechanistic quantities carry an identifiability audit with explicit abstention rates. For this candidate the verdicts are succinct: active but not beneficial, causal but not superior, nominally calibrated but uninformative, and rarely and unstably identifiable. Future priorities include pressure-range-aware objectives, sharper shift-sensitive uncertainty, target-independent timing control, and evaluation on ambulatory recordings.

Subject of Research: Stress-testing a physics-constrained arterial-pressure digital twin candidate for cuffless blood pressure reconstruction and forecasting from photoplethysmography

Article Title: Stress testing a physics-constrained arterial-pressure twin candidate: Reconstruction, forecasting, uncertainty, and identifiability from photoplethysmography

Article References: Hussein, A. F. (2026). Stress testing a physics-constrained arterial-pressure twin candidate: Reconstruction, forecasting, uncertainty, and identifiability from photoplethysmography. Results in Engineering, 32, Article 113759. https://doi.org/10.1016/j.rineng.2026.113759

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113759

Keywords: digital twin, photoplethysmography, blood pressure, Windkessel model, physics-informed learning, uncertainty quantification, conformal prediction, identifiability, PulseDB, cuffless monitoring, machine learning, cardiovascular modeling