Non-Invasive BP Estimation via Carotid-Brachial-Radial Coupling: A Physics-Informed Asymmetric Cycle Network Approach

Photoplethysmogram (PPG) based cuffless blood pressure (BP) monitoring holds significant potential to revolutionize hypertension management. However, its widespread adoption is hampered by limited accuracy and poor interpretability, primarily due to the lack of central arterial information and physical constraints. To address the inability of PPG to capture central hemodynamic drivers, we proposed a novel multimodal approach that leveraged multi-site physiological coupling from the carotid, brachial, and radial arteries, integrated into a physics-informed deep learning framework called the PIAC-Net. Our method integrated carotid blood flow velocity - measured using an ultrasound sensor - as a key central physiological signal, thereby supplementing the peripheral information provided by PPG alone. The PIAC-Net incorporated a set of hybrid physical loss functions that integrated 1D Navier-Stokes equations with 0D Windkessel boundary conditions, thereby enforcing strict hemodynamic conservation constraints on the latent brachial waveform used as the BP reference. Five-fold cross-validation showed superior accuracy than compared methods. The mean absolute errors of systolic/diastolic BP on the synthetic (N = 4,374) and in-house (N = 1,765) datasets were 1.82/1.51 mmHg and 7.08/6.38 mmHg, respectively. Ablation studies demonstrated that integrating carotid information reduced estimation errors by over 50%, serving as the primary accuracy factor, while the physics-based constraint mechanism also exhibited consistent high-performance estimation under small-sample conditions. This dual-fusion paradigm of "central driver-peripheral response" and "data-physics" integration offers a clinically viable solution for high-precision and interpretable BP monitoring.