PPG-Based Feature Extraction for Type II Diabetes Prediction: A Machine Learning Approach

The timely and precise diagnosis of diabetes holds paramount importance for efficient management and prevention of associated complications. While conventional diagnostic methodologies remain reliable, they are often invasive, resource-intensive, and time-consuming. This study endeavors to present a non-invasive methodology for diabetes detection through the utilization of Photoplethysmogram (PPG) signals. A diverse array of features is meticulously extracted from the PPG data, encompassing Spectral entropy, Shannon energy, and various time-frequency domain features. The importance of each feature in diabetes prediction is assessed through machine learning techniques like Random Forest and extreme gradient boosting. Employing these features, the PPG signals are effectively classified into diabetes (type II) or normal categories using multiple machine-learning models. Through extensive experimentation, it is found that the proposed method outperforms the recent studies in a publicly available dataset. The results demonstrate high accuracy and robustness across different feature subsets, showcasing substantial promise for real-world applications in non-invasive diabetes detection.