Toward Non-Invasive Calibration-Free Blood Pressure Estimation: A Physics-Informed Neural Network with Hemodynamic Priors

Continuous and calibration-free arterial blood pressure (ABP) monitoring remains a major challenge in wearable sensing. Most data-driven estimators rely on ECG–PPG correlations without enforcing the underlying cardiovascular physics, which limits interpretability and cross-subject generalization. We introduce a hemodynamics-driven Physics-Informed Neural Network (PINN) that reconstructs the full radial pressure waveform from ECG and fingertip PPG. The model embeds a 1D pulse-wave propagation PDE, Windkessel boundary conditions, and spatially varying vessel geometry and physiology, treated as soft physiological constraints. Trained on MIMIC III database and evaluated on the large clinical database VitalDB, the model achieved an inter-subject MAE of 13.10mmHg, with SBP 15.84mmHg and DBP 10.42mmHg. Beyond waveform reconstruction, the framework predicts also additional hidden physiological quantities, such as radius and wall thickness, elasticity, flow and spatial pressure evolution, consistently with the governing equations. These results highlight both the promise and the current limitations of physics-informed modeling, positioning this work as a foundational step toward future personalized, physiologically grounded, and calibration-free cuffless BP monitoring, while remaining not ISO 81060-2 compliant or clinically deployable in its current form.