Physiology-Informed Neural Network Cuts Training Data for Cuffless Blood Pressure Waveforms

Blood pressure is not a single number but a continuously undulating waveform, rising and falling with every heartbeat as the heart contracts and the arteries recoil. Capturing that full waveform without a cuff has long been a goal of wearable health technology, and a new proof-of-concept study published in Artery Research reports a step toward it. Researchers Zaineb Aloui and Cederick Landry describe a dual-output, physiology-informed neural network that estimates the complete blood pressure waveform from electrocardiography (ECG) and photoplethysmography (PPG) signals while requiring only fifteen minutes of subject-specific training data—half the amount its predecessor needed.

The work builds on a nonlinear autoregressive model with exogenous inputs, or NARX, a neural network architecture that predicts blood pressure one time step ahead using recent samples of PPG and ECG along with its own previous blood pressure estimates. NARX models trained on individual subjects had previously shown promise for reconstructing the full waveform during daily life, but they demanded complex training data: two fifteen-minute calibration segments, one at the start and one at the end of the monitoring period, deliberately including physical activity to span a wide range of blood pressure values. That burden, the authors argue, limits clinical adoption.

The new architecture, called NARXphysio, attacks the data problem by embedding cardiovascular physiology directly into the network. It pairs the standard NARX blood pressure estimator with a second sub-model: a parametric sigmoidal layer with four trainable parameters that converts the estimated blood pressure waveform into an estimated PPG signal. The sigmoid reflects a well-known relationship in vascular physiology—as blood pressure rises, arteries dilate, blood volume within them increases, and the PPG sensor registers that volume change in a nonlinear, sigmoid-shaped fashion. During training, both outputs feed a shared composite loss function combining mean squared and mean absolute errors for blood pressure with reconstruction errors for the PPG waveform and its first derivative, so the physiological constraint directly shapes how the network learns.

The evaluation rested on a distinctive dataset. Five healthy adults, four men with an average age of 28 years, free of cardiovascular and peripheral vascular disease, wore an Astroskin vest recording forehead PPG and ECG alongside Finapres finger-cuff blood pressure waveforms at 64 Hz for roughly six and a half hours. The protocol included two fifteen-minute standardized sequences featuring postures, static handgrip, Valsalva maneuvers and walking, plus six hours of unconstrained activities of daily living. Because continuous blood pressure waveforms during real-world activity are rarely captured—the Portapres device used is now discontinued—the authors note that no comparable public dataset exists, even though the sample is small.

Three model configurations were compared on the six hours of unseen daily-activity data: a baseline NARX trained on the full thirty minutes, a reduced-data NARX trained on only the first fifteen minutes, and NARXphysio, also trained on fifteen minutes. A fourth benchmark, the mean training data model, simply predicted a constant systolic and diastolic value derived from the first fifteen minutes, simulating what a single morning cuff measurement could offer. Performance was scored beat by beat using mean absolute error, mean error, error standard deviation and Pearson correlation, averaged across participants and compared with Wilcoxon rank sum tests and Bonferroni correction.

The headline result is that NARXphysio, trained on half the data, matched or exceeded the thirty-minute baseline. Its mean absolute error was 4.92 ± 5.7 mmHg for systolic and 4.4 ± 4.96 mmHg for diastolic pressure, against 5.77 ± 6.59 and 4.58 ± 5.54 mmHg for the thirty-minute NARX. The reduced-data NARX without the physiological layer fared worse, and ablation studies confirmed that removing either the PPG reconstruction term or its derivative term from the loss degraded performance, raising systolic errors to between roughly 6.5 and 6.7 mmHg. NARXphysio was the only model to show a nearly significant improvement in mean absolute error over the constant-value benchmark (p = 0.009); no other between-model difference reached significance, which the authors attribute to the sample of five.

Qualitatively, the model tracked systolic peaks, diastolic troughs and the dicrotic notch with errors as low as 1.4 to 3.0 mmHg in representative segments, compared with 4.2 to 6.2 mmHg for the reduced-data baseline, and it followed rapid blood pressure swings of up to 30 mmHg within a few heartbeats. During Valsalva maneuvers it tended to underestimate dicrotic notches, a limitation the authors link to peripheral vasoconstriction altering the arterial wall’s viscoelastic properties. Error trajectories over six hours, smoothed with twenty-minute moving averages, remained bounded for all models, but the fifteen-minute baseline drifted toward a negative bias while NARXphysio stayed centered near zero, suggesting better temporal stability.

Beyond accuracy, the sigmoidal layer yields interpretable physiological parameters unavailable in conventional black-box networks. The width of the estimated arterial compliance curve, quantified as the full width at half maximum of the sigmoid’s derivative, averaged 36.5 mmHg—close to the roughly 33.4 mmHg obtained by reanalyzing an earlier finger-artery model, which the authors take as support for the sigmoid’s physiological relevance. The contact pressure parameter averaged 99 mmHg at the forehead sensor site; correcting for the roughly 40 mmHg hydrostatic difference between forehead and heart level implies an average PPG contact pressure of 59 mmHg, consistent with acceptable signal quality and comfort. These parameters showed substantial inter-individual variability, underscoring the personalized nature of the model.

Performance varied with activity. Both error magnitude and variability increased from sitting to standing to walking, with sit-to-stand transitions—rapid blood pressure events concentrated in a small fraction of the data—contributing disproportionately. NARXphysio showed reduced variability across participants during sitting, and its advantages were more consistent for systolic than diastolic pressure. The authors also note that modest correlation coefficients for systolic estimates (0.43–0.72) partly reflect the limited natural blood pressure variability of daily life relative to measurement noise, and that simpler pulse transit time models on the same dataset yielded correlations of only about 0.2 to 0.4.

The limitations are candid. Five participants cannot represent the physiological diversity of broader populations, and conditions such as intense exertion, acute hypotension and nighttime blood pressure remain untested. Fifteen minutes of calibration is still considerable, and reducing it further would require protocols that provoke a wider range of hemodynamic changes. Variations in PPG contact pressure during daily activities may masquerade as blood pressure changes, motivating future sensors that measure contact pressure directly. Still, the authors envision a practical pathway: a patient begins a twenty-four-hour monitoring session with a fifteen-minute calibration on a clinic volume-clamping device, then continues daily life wearing seamless ECG and PPG sensors. Scaling from five people over six hours to that vision, they caution, will require substantial further work—but this pilot demonstrates that physiological structure, built into the network itself, can halve the data a personalized cuffless blood pressure model needs.

Subject of Research: A physiology-informed neural network for subject-specific cuffless blood pressure waveform estimation

Article Title: A Dual-Output Physiology-Informed Neural Network Architecture for Continuous Cuffless Blood Pressure Waveform Estimation: A Proof‑of‑Concept

Article References: Aloui, Z., & Landry, C. (2026). A Dual-Output Physiology-Informed Neural Network Architecture for Continuous Cuffless Blood Pressure Waveform Estimation: A Proof‑of‑Concept. Artery Research, 32(1), Article 21. https://doi.org/10.1007/s44200-026-00120-3

Image Credits: AI Generated

DOI: 10.1007/s44200-026-00120-3

Keywords: cuffless blood pressure, photoplethysmography, NARX neural network, physiology-informed machine learning, BP waveform, wearable sensors, pulse wave analysis, arterial compliance, ECG, personalized calibration, hypertension monitoring, proof of concept