A Feasibility Study on Noninvasive Blood Glucose Estimation Using Machine Learning Analysis of Near-Infrared Spectroscopy Data

This study explored the feasibility of noninvasive blood glucose (BG) estimation using near-infrared (NIR) spectroscopy with dog blood samples. A sensor module employing three representative wavelengths (770 nm, 850 nm, and 970 nm) was tested on an artificial blood vessel (ABV) and a thin pig skin (TPS) model. BG concentrations were adjusted through dilution and enrichment with injection-grade water and glucose solution, and reference values were obtained from three commercial invasive glucometers. Correlations between NIR spectral responses and glucose variations were quantitatively evaluated using linear, multiple, partial least squares (PLS), logistic regression, regularized linear models, and multilayer perceptron (MLP) analysis. The results revealed distinct negative correlations at 850 nm and 970 nm, identifying these wavelengths as promising candidates for noninvasive glucose sensing. Furthermore, an NIR-glucose database generated from actual dog blood was established, which may serve as a valuable resource for the development of future noninvasive glucose monitoring systems.