Noninvasive continuous blood pressure prediction using FlexNIRS and machine learning during carotid endarterectomy

Continuous blood pressure (BP) monitoring is crucial for maintaining hemodynamic stability and complication prevention. Near-infrared spectroscopy photoplethysmography (NIRS-PPG) offers a noninvasive alternative to arterial lines (A-line) for continuous BP monitoring.

We aim to assess whether a wearable NIRS-PPG device (FlexNIRS) can estimate mean arterial pressure (MAP) using linear and Gaussian process regression (GPR) models.

NIRS-PPG signals were recorded bilaterally in 10 patients undergoing carotid endarterectomy. Subject-specific linear regression and GPR models predicted MAP based on heart rate and peak features of the NIRS-PPG signal. A-line readings served as the reference.

All models achieved strong performance with $R^2\ge 0.75$ . The two-feature GPR model improved accuracy ( $R^2=0.78$ ), whereas adding a third feature further enhanced performance ( $R^2=0.82$ ). Improvements in $R^2$ , mean absolute error, and root mean squared error were statistically significant. The highest accuracy was observed contralateral to the surgical site using the 2.8-cm source-detector separation.

This preliminary study supports the feasibility of noninvasive MAP estimation using NIRS-PPG and machine learning. The approach may provide a practical alternative for BP monitoring after A-line removal in postoperative and intensive care unit settings.