Optimal Wavelength Bands for Remote Blood Glucose Estimation Using Facial NIR Spectroscopic Images Measured At 800–1650 Nm

Traditional blood glucose measurement methods are invasive, leading to discomfort and inconvenience. Recently, non-invasive techniques using near-infrared (NIR) spectroscopy have gained attention. This study explores a novel approach to estimating blood glucose levels by analyzing spatial features from facial NIR images. Unlike previous studies that have relied on broadband NIR light sources (760-1650 nm), this study focuses on identifying an optimal narrowband wavelength to reduce noise and enable simpler hardware implementation. Facial NIR spectroscopic images were captured using discrete light sources spanning 800-1650 nm, and spatial features were extracted using independent component analysis (ICA) from single-frame images. The estimation accuracy was maximized at 1200 nm, with spatial features strongly expressed along the side of the nose-corresponding to major superficial arteries. Although the root mean square error ($RMSE$) was higher than that of contact-based systems, the method offers a favorable trade-off for unobtrusive, scalable monitoring. The analysis identified 1200 nm as a practically viable wavelength for non-contact blood glucose estimation. These findings support the feasibility of a compact, low-cost NIR imaging system for real-world deployment in non-invasive glucose monitoring applications such as telehealth, smart homes, or public health settings.