PMB-NN: Physiology-Centred Hybrid AI for Personalized Hemodynamic Monitoring from Photoplethysmography

Continuous monitoring of blood pressure (BP) and hemodynamic parameters such as peripheral resistance (R) and arterial compliance (C) are critical for early vascular dysfunction detection. While photoplethysmography (PPG) wearables has gained popularity, existing data-driven methods for BP estimation lack interpretability. We advanced our previously proposed physiology-centered hybrid AI method-Physiological Model-Based Neural Network (PMB-NN)-in blood pressure estimation, that unifies deep learning with a 2-element Windkessel based model parameterized by R and C acting as physics constraints. The PMB-NN model was trained in a subject-specific manner using PPG-derived timing features, while demographic information was used to infer an intermediate variable: cardiac output. We validated the model on 10 healthy adults performing static and cycling activities across two days for model’s day-to-day robustness, benchmarked against deep learning (DL) models (FCNN, CNN-LSTM, Transformer) and standalone Windkessel based physiological model (PM). Validation was conducted on three perspectives: accuracy, interpretability and plausibility. PMB-NN achieved systolic BP accuracy (MAE: 7.2 mmHg) comparable to DL benchmarks, diastolic performance (MAE: 3.9 mmHg) lower than DL models. However, PMB-NN exhibited higher physiological plausibility than both DL baselines and PM, suggesting that the hybrid architecture unifies and enhances the respective merits of physiological principles and data-driven techniques. Beyond BP, PMB-NN identified R (ME: 0.15 mmHg$\cdot$s/ml) and C (ME: -0.35 ml/mmHg) during training with accuracy similar to PM, demonstrating that the embedded physiological constraints confer interpretability to the hybrid AI framework. These results position PMB-NN as a balanced, physiologically grounded alternative to purely data-driven approaches for daily hemodynamic monitoring.

💡 Research Summary

This paper introduces a physiology‑centred hybrid artificial intelligence framework, the Physiological Model‑Based Neural Network (PMB‑NN), for continuous, personalized hemodynamic monitoring using photoplethysmography (PPG). The authors aim to overcome the interpretability limitations of purely data‑driven blood pressure (BP) estimators while retaining their robustness to noisy wearable signals. To this end, PMB‑NN integrates a two‑element Windkessel model—parameterized by total peripheral resistance (R) and arterial compliance (C)—as physics‑based constraints within the loss function of a feed‑forward neural network.

Data were collected from ten healthy adults (mixed gender, ages 22‑35) over two separate days. Each session comprised three activity phases: static posture (seated then standing), low‑intensity cycling (45 rpm, 50 W), and moderate‑intensity cycling (45 rpm, 100 W). PPG was recorded at 64 Hz from the index finger using a Shimmer sensor, while continuous arterial pressure was obtained via a Finometer volume‑clamp finger cuff. The authors extracted beat‑to‑beat timing features—systolic upstroke time (Ts) and diastolic time (Td)—from the PPG waveform, and derived reference values for systolic/diastolic BP, cardiac output (Q), R, and C from the pressure waveform using established algorithms.

A separate fully‑connected “Q‑network” was trained in a leave‑one‑subject‑out (LOSO) fashion to estimate cardiac output from Ts, Td, and demographic variables (age, sex, height, weight). The optimal Q‑network architecture (6‑64‑128‑64‑1) achieved low mean‑squared error and was then applied to the second‑day data to provide subject‑specific Q estimates for PMB‑NN input.

PMB‑NN receives a concatenated sequence of Ts and Td values together with the estimated Q. Its core consists of three hidden layers (128 units each) with ReLU activations. The loss combines the standard mean‑squared error between predicted and reference systolic/diastolic pressures and a physics‑based penalty enforcing the Windkessel relationship (P = R·Q + C·dP/dt). This coupling forces the network’s outputs to obey hemodynamic principles, thereby yielding interpretable estimates of R and C alongside BP.

The model was benchmarked against three state‑of‑the‑art deep learning baselines—fully connected neural network (FCNN), CNN‑LSTM, and Transformer—and against a standalone Windkessel physiological model (PM). Evaluation considered three dimensions: (1) accuracy (MAE, MSE), (2) physiological interpretability (whether inferred R and C remain within plausible ranges and follow expected trends across activity levels), and (3) physiological plausibility (a composite metric reflecting consistency with known cardiovascular physiology).

Results showed that PMB‑NN achieved a median systolic BP MAE of 7.2 mmHg, comparable to the deep learning baselines, while its diastolic MAE was 3.9 mmHg, slightly higher than the DL models. Importantly, PMB‑NN demonstrated markedly superior physiological plausibility: inferred R and C values were physiologically realistic, exhibited the expected increase in resistance and decrease in compliance during exercise, and remained stable across the two testing days, indicating good day‑to‑day robustness. The standalone physiological model performed poorly in BP accuracy, and the pure DL models, although accurate, produced R and C estimates that were inconsistent with known physiology.

The authors conclude that embedding a simple Windkessel model as a soft constraint within a neural network yields a balanced solution: accurate BP estimation, simultaneous recovery of key hemodynamic parameters, and enhanced interpretability. This hybrid approach is positioned as a viable alternative for wearable‑based continuous monitoring, especially in clinical scenarios where trust and explainability are essential. Future work is suggested to extend validation to hypertensive and elderly populations, incorporate multimodal sensor data (e.g., accelerometry, multi‑wavelength PPG), and explore model compression for real‑time deployment on low‑power wearable platforms.