Machine Learning for Classifying Vascular Age Using the Second Derivative of the PPG Signal

Abstract

The second derivative of the photoplethysmographic signal features characteristic peaks that provide information on vascular properties. Our analysis was conducted on deflection wave features of the second derivate of PPG signals from human volunteers aged between 26 to 63 years. We show that machine learning models can leverage these features to accurately distinguish subjects based on their age group in a binary classification approach, achieving a balanced accuracy of 80% and an AUC of 0.88 on unseen data, and 100% accuracy on a sample-level split. When including the ratios of deflection wave peaks in our feature set, accuracies of up to 89% and AUC of 0.95 were achieved. Our feature plot illustrates clear clustering between younger and older volunteers, specifically when using the peak ratios, confirming its discriminatory power and its significant potential as a non-invasive index for assessing vascular age. This study demonstrates that second derivative PPG peak ratios, together with the peak amplitudes, are a more robust marker of vascular age than the peak amplitudes alone.