Non-invasive assessment of pulmonary capillary wedge pressure using a dual-modal wearable system and physics-informed artificial intelligence: a validation study against right heart catheterization
T Zou, T Zou, J Wang, Q W Xiong, Q W Xiong, D H Liu, D H Liu, Y Zou, Y ZouAbstract
Background
Hemodynamic congestion, characterized by elevated filling pressures, is a primary driver of rehospitalization and poor prognosis in heart failure (HF) patients. Although pulmonary capillary wedge pressure (PCWP) measured via right heart catheterization (RHC) remains the gold standard for assessing congestion , its invasive nature precludes continuous monitoring in the home setting. This study aims to validate a novel dual-modal wearable platform and a physics-informed algorithm for medical-grade, non-invasive PCWP assessment.
Methods
A comprehensive monitoring system was developed, comprising a biomimetic flexible piezoelectric patch for capturing cardiomechanical signals (seismocardiography, SCG; phonocardiography, PCG) and a wrist-worn device for photoplethysmography (PPG). The patch utilizes a PVDF/ZnO composite with bionic micro-protrusions to optimize skin acoustic impedance matching. Using a high-precision time synchronization mechanism (<2 ms), key clinical features including pre-ejection period (PEP), pulse transit time (PTT), and diastolic heart sounds (S3 and S4) were extracted. A Physics-Informed Neural Network (PINN) was employed to estimate PCWP by embedding hemodynamic equations (e.g., the Bramwell-Hill equation) into the deep learning loss function, ensuring physical consistency and enhancing accuracy in data-scarce medical scenarios. Individualized model calibration was achieved through a "digital twin" approach using synchronized in-hospital RHC data.
Results
The biomimetic sensor design improved the signal-to-noise ratio (SNR) by over 30% compared to conventional flat sensors, facilitating the reliable detection of faint S3 and S4 signals indicative of HF decompensation. A hardware-level differential noise-reduction structure effectively suppressed motion artifacts during daily activities, ensuring data continuity. Clinical validation against RHC showed that the PINN-based non-invasive estimates were highly correlated with gold-standard PCWP values. Notably, the physics-informed constraint corrected prediction errors that violated physiological norms, significantly outperforming pure "black-box" deep learning models.
Conclusion
By integrating biomimetic flexible sensing with physics-guided AI, this platform enables the transition from qualitative trend monitoring to quantitative, medical-grade PCWP assessment. This "in-hospital calibration, out-of-hospital tracking" paradigm provides a robust tool for remote hemodynamic management, potentially reducing rehospitalization through early intervention in heart failure patients.Biomimetic Dual-modal Sensing SystemFor image description, please refer to the figure legend and surrounding text.Clinical Gold-standard ValidationFor image description, please refer to the figure legend and surrounding text.