Non-invasive prediction of cholesterol levels from photoplethysmogram (PPG)-based features using machine learning techniques: a proof-of-concept study

Regular monitoring of cholesterol levels is crucial to reducing the risk of vascular blockage and preventing atherosclerotic cardiovascular diseases. However, standardized cholesterol measurement tests involve 8-12 hours of strict fasting and blood extraction via finger pricking, which may cause pain and discomfort. This study explores the usefulness of fiducial-based features extracted from the photoplethysmogram (PPG) in estimating blood cholesterol by combining feature selection methods and machine learning techniques. We extracted 150 features from forty-six 2-minute PPG recordings and included participants' age as a feature. Several variations of linear regressions (LR), regression trees (RT), support vector regressions (SVR), and Gaussian process regressions (GPR) were trained with the most relevant features. The rational quadratic GPR model achieved the lowest errors (MAE = 11.70, MSE = 281.57, and RMSE = 16.78 mg/dL) and the highest coefficient of determination (r(2) = 0.832) when combined with ReliefF. The proposed method holds promise for developing lightweight and non-invasive approaches for blood cholesterol estimation, although it may require further validation due to the limited sample size.