Machine Learning-Based Diabetes Detection Using Photoplethysmography Signal Features

Abstract

Diabetes is a chronic condition which prevention and control is done mostly by minimally invasive devices. In this work, we propose a noninvasive method based on photoplethysmography (PPG) for cost-effective and discomfort-free diabetes detection and prevention. We used PPG signal features and patient metadata from a public dataset for classifying subjects as Diabetic or non-Diabetic. The Logistic Regression and eXtreme Gradient Boosting algorithms were evaluated using a five-fold cross validation approach and achieved a mean AUC of 0.79 ± 0.15 and 0.73 ± 0.17, respectively. Our results align with existing literature, supporting the use of machine learning techniques for developing non-invasive diabetes detection and prevention devices.