Noninvasive Blood Glucose Monitoring with Machine Learning Enhanced Transmittance Spectroscopy

Diabetes is a widespread metabolic disease in which the body's inability to regulate blood glucose levels leads to severe health complications. Current limitations in noninvasive glucose sensing make finger-prick glucometers the standard for personal monitoring despite their discomfort and challenges for frequent measurements. By combining machine learning techniques with transmittance spectroscopy, this study presents a noninvasive approach for estimating BGL in personal healthcare. The system employs multispectral transmittance measurements at 650, 808, and 940 nm to evaluate glucose concentrations in aqueous solutions. Simulation of light absorption in skin layers and in vitro experiments identified 940 nm as the optimal wavelength, offering high sensitivity with minimal interference from water absorption. Using this wavelength, an in vivo system based on transmittance spectroscopy was developed for noninvasive glucose measurement. A machine learning pipeline incorporating ensemble models was implemented to predict glucose levels from optical data. Trained on 200 clinical samples obtained from the in vivo experimental setup, the XGBoost model outperformed other algorithms, achieving an R ² score of 0.94 with a root-mean-squared error (RMSE) of 23.92 mg/dL. The Clarke error grid analysis (CEGA) showed that 92.5% of predictions fell within Zone A and 7.5% within Zone B, with no samples in the critical zones (C-E), demonstrating the system's robustness. These results highlight the potential of the proposed method as a practical, low-cost, and reliable solution for continuous noninvasive glucose monitoring in diabetes management.