Enhanced Non-Invasive Diabetes Classification with Effective Photoplethysmogram Signal Features for Future Point-of-Care Diagnostics

The global diabetes point-of-care (POC) biosensor market has seen substantial growth, driven by the increasing prevalence of diabetes and the demand for non-invasive glucose monitoring technologies. Photoplethysmogram (PPG) signal analysis has emerged as a promising method for non-invasive diabetes detection, offering a cost-effective and user-friendly alternative to traditional glucose monitoring systems. This study explores the potential of PPG signal features for enhancing the accuracy of diabetes classification using machine learning models. By focusing on key features such as Mean Inter-Beat Interval (IBI), Standard Deviation of NN Intervals (SDNN), Root Mean Square of Successive Differences (RMSSD), Length Max to Rastio, and Spectral Entropy, the study aims to improve the reliability of non-invasive diabetes detection methods. The study evaluated three machine learning models: LightGBM (LGBM), Random Forest (RF), and Support Vector Machine (SVM). The results demonstrate that LGBM achieved the highest accuracy of 98.86%, followed by RF with 96.59%, and SVM with 89.77%. These findings suggest that PPG signal analysis, combined with robust machine learning techniques, holds significant potential for advancing non-invasive diabetes monitoring technologies, offering a more accessible and effective solution for early detection and continuous monitoring.