Exploring supervised machine learning models to estimate blood pressure using non-fiducial features of the photoplethysmogram (PPG) and its derivatives

Machine learning has proven valuable in developing photoplethysmography (PPG)-based approaches for blood pressure (BP) estimation, with many holding some promise for cuff-less BP assessment. Still, their efficacy relies on accurate and robust fiducial point detection algorithms. The present study explores the usefulness of several non-fiducial features of the PPG signal and its derivatives in estimating BP by combining well-known feature selection methods and machine learning techniques. We collected PPG recordings from 56 participants and computed fifty-seven non-fiducial features, including statistical indexes and energy operators. After implementing three feature selection algorithms (i.e., F-test, mRMR, and ReliefF), the most relevant features were employed to train four learning regression model families. We computed the mean of all absolute errors (MAE), the squared sum and the standard deviation of the errors (MSE and RMSE, respectively), and the coefficient of determination (r²) to evaluate the performance of each model. The abovementioned feature selection methods produced different optimal feature subsets for systolic and diastolic BP values, with the Matern 5/2 and the rational quadratic GPR models providing the best predictions when combined with ReliefF (MAE = 0.44, MSE = 0.61, and RMSE = 0.78 mmHg for SBP; MAE = 0.31, MSE = 0.40, and RMSE = 0.63 mmHg for DBP). Furthermore, each model utilizes only fifteen easy-to-compute features, thus becoming suitable for computationally constrained hardware. We highlight the need for implementing feature selection algorithms exhaustively, as the most relevant PPG-based features for systolic and diastolic BP estimation might not have the same weight.