PMB-NN: Physiology-centered 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 and therapy optimization. While photoplethysmography (PPG) wearables has gained widespread popularity, existing data-driven methods for BP estimation lack physiological interpretability.

We advanced our previously proposed physiology-centered hybrid AI method-the Physiological Model-Based Neural Network (PMB-NN)-in arterial hemodynamic monitoring, that unifies deep learning with a 2-element Windkessel based physiological model (PM) 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 as inputs, while demographic information was used to infer cardiac output (Q) which is integrated as a prerequisite to solve PM for R and C during training. The model outputs personalized systolic and diastolic BP. The PMB-NN was initially validated on a primary cohort of 10 healthy young adults performing multi-day static and cycling activities to assess day-to-day robustness, benchmarked against established deep learning (DL) models (FCNN, CNN-LSTM, and Transformer) as well as the standalone PM. This core evaluation followed a tripartite framework: (i) estimation accuracy for BP; (ii) physiologically constrained interpretability, reflected by the model's ability to infer R and C; and (iii) physiological plausibility, assessed via the correlation between estimated BP and input timing features. To explore the mechanistic and operational boundaries of the framework, we quantified the impacts of motion artifacts, Q estimation errors, and calibration efficiency, while assessing model transferability from exercise-driven steady states to distinct autonomic challenges across an age-diverse cohort (n=37, 18-88 years old).

PMB-NN achieved systolic BP accuracy (median MAE: 7.2 mmHg) comparable to DL benchmarks, despite yielding diastolic performance (median MAE: 3.9 mmHg) lower than DL models. Crucially, however, PMB-NN exhibited substantially 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 also identified R (median MAE: 0.13 mmHg s/ml) and C (median MAE: 0.19 ml/mmHg) during training, achieving parameter estimation fidelity similar to the standalone PM, demonstrating that the embedded physiological constraints confer interpretability to the hybrid AI framework. Furthermore, extended evaluations confirmed the framework's resilience to moderate motion artifacts (SNR = 15 dB) and identified 5-min calibration as a pragmatic threshold for personalization. While numerical errors increased under autonomic challenges, it captured age-related hemodynamic shifts while maintaining consistent physiological plausibility across the adult lifespan.

These results position PMB-NN as a balanced, physiologically grounded alternative to purely data-driven approaches for daily hemodynamic monitoring.