Cuffless Blood Pressure Estimation From Wearable Photoplethysmography and Windkessel Model-Regularized Learning Approach

Abstract

Wearable cuffless blood pressure (BP) monitoring using photoplethysmography (PPG) sensors has gained attention for noninvasive hemodynamic assessment. However, most deep learning methods are purely data-driven and lack physiological consistency. This study presents a physiology-informed learning framework for PPG-based BP estimation by integrating cardiovascular dynamics derived from the Windkessel model into a bounded loss function. The proposed framework incorporates physiological constraints by enforcing exponentially decaying arterial pressure using a PPG-derived diastolic time constant. PPG signals from ten healthy subjects were collected using a wearable finger-based sensor during rest, affective stimulation, and Stroop test conditions to introduce variability. The system complies with Association for the Advancement of Medical Instrumentation standards, achieving a mean absolute error of 2.90 <inline-formula><tex-math notation="LaTeX">$\pm$</tex-math></inline-formula> 1.20 mmHg for DBP and 4.10 <inline-formula><tex-math notation="LaTeX">$\pm$</tex-math></inline-formula> 1.41 mmHg for SBP. These results, from a small, homogeneous cohort, support feasibility of physiology-informed learning and motivate validation on larger, diverse populations toward clinical translation.