Evaluation of Blood Glucose Estimation Models for Noninvasive Glucose Monitoring System Using Visible and Infrared Spectroscopy
TL;DR
A dual-wavelength transillumination measurement system using 660- and 940-nm light and evaluates multiple regression models for noninvasive blood glucose estimation met the <100-mg/dL threshold and approached the $\geq 100$ -mg/dL threshold.
Abstract
Blood glucose monitoring is essential for metabolic health management, yet noninvasive optical estimation remains challenging because glucose-related optical signatures are weak and are easily confounded by tissue heterogeneity and hemodynamic variation. This study presents a dual-wavelength transillumination measurement system using 660- and 940-nm light and evaluates multiple regression models for noninvasive blood glucose estimation. The 660-nm channel serves as a physiological reference, whereas the 940-nm channel provides glucose-sensitive near-infrared (NIR) information. In a two-phase study, models were developed on measurements from one cohort and then applied without retraining to a separate validation cohort, using finger-stick readings from a commercial invasive glucometer as the comparative reference. Among the evaluated models, a feedforward neural network (FNN) achieved the best performance on the validation set, with a mean absolute error (MAE) of 7.713 mg/dL, a root-mean-square error of 9.758 mg/dL, a mean absolute relative difference of 6.927 %, and an $R^{2}$ of 0.865. The dual-wavelength configuration outperformed single-wavelength models and an E500 smartwatch benchmark. The Clarke error grid analysis placed 98.8 % of estimates in Zone A and 1.2 % in Zone B. Relative to ISO 15197 accuracy criteria, the proposed system met the <100-mg/dL threshold and approached the $\geq 100$ -mg/dL threshold.