A Dual-Output Physiology-Informed Neural Network Architecture for Continuous Cuffless Blood Pressure Waveform...

This pilot study introduces the first subject‑specific, physiology‑informed neural network approach for accurate estimation of the full blood pressure (BP) waveform. The method builds on a nonlinear autoregressive model with exogenous inputs (NARX), implemented through artificial neural networks trained on subject‑specific electrocardiography (ECG) and photoplethysmography (PPG) signals. Three model configurations are evaluated: (1) a reference NARX model trained on 30 min of data, (2) a reduced‑data NARX model trained on 15 min, and (3) a physiology‑informed model (NARX_physio), also trained on 15 min, which incorporates a personalized four‑parameter BP-PPG sigmoidal layer into the training process through a dual‑output loss function that simultaneously estimates the BP and PPG waveforms. By embedding BP-PPG physiology directly into training, the architecture ensures internal physiological consistency-something not achieved in existing PPG‑based or purely data‑driven machine learning approaches. All three models were assessed on a small (n = 5), but rich dataset, covering over six hours of activities of daily living. Despite using only half the training duration of the baseline NARX model, the physiology‑informed approach maintains accurate BP estimation. Beyond improved accuracy, the sigmoidal layer provides physiologically interpretable parameters, including estimated arterial‑compliance curve width with an average of 36.5 mmHg and an estimated PPG contact pressure with an average of 59 mmHg in this study, offering insights unavailable in standard neural network models. Together, these preliminary results suggest that the NARX_physio model enables effective personalization of cuffless BP waveform estimation with limited training data, highlighting its potential to improve both the reliability and the clinical relevance of continuous cuffless BP monitoring.