Machine Learning Techniques Applied to in-Silico Pulse Wave Velocity Estimation Based on Photoplethysmographic Signals

Carotid femoral pulse wave velocity (cf-PWV) is the gold standard measurement of arterial stiffness that has been recognized as an effective cardiovascular disease (CVD) risk biomarker. Although reliable and accurate, the reference method for measuring cf-PWV is time-consuming and requires the intervention of a qualified practitioner. Photoplethysmography (PPG) is a non-invasive cost-effective technology that contains a variety of information about the cardiovascular system. This paper aims to explore the potential of estimating cf-PWV through PPG pulse wave analysis for large-scale CVD risk screening. Our work includes a comparative analysis involving two machine learning models and various sensor positions. A set of features based on fiducial points extracted from in-silico PPG signals and their derivatives is used as an input for XGBoost and Support Vector Regression (SVR) models. These models were trained on simulated sensor positions, evaluated across different noise levels, and demonstrated comparable or superior performance compared to previous studies. The proposed method is deployable on a low-power embedded processor. Signals from the superficial temporal artery position exhibit the best performance, achieving an R2 of 1.00 and a root-mean-square error (RMSE) of 0.13. The PPG signal combined with the proposed method shows promising potential for cf-PWV estimation particularly when using superficial temporal artery signals. Our results paved the way for future in-vivo validation of the described method.