Generalization Performance Evaluation of a Blood Glucose Estimation Model Based on Near-Infrared Facial Images with Wavelengths Ranging from 760 nm to 1650 nm

抄録

With the aim of realizing remote blood glucose level measurement, we constructed a model for estimating blood glucose level base on facial images measured in the near-infrared band, which is highly transparent to living tissue. However, the generalization performance of the estimation model was not evaluated in previous studies due to the small number of data. The objective of this study is to construct an individual model and a general model for blood glucose estimation based on facial images measured in the near-infrared broadband wavelength range of 760 nm to 1650 nm, and to evaluate their generalization performance. Independent Component Analysis(ICA) was applied to facial images during blood glucose variations to obtain spatial features of independent components and their weights. Using the standardized weights as explanatory variables and reference blood glucose levels as objective variables, an individual model and a general model for blood glucose estimation were constructed through multiple regression analysis. Cross-validation was applied to evaluate the generalization performance of the models. The results showed that the accuracies of blood glucose estimation in the 760nm to 1100 nm and 1050nm to 1650 nm wavelength bands were 32.23 mg/dL and 36.10 mg/dL in RMSE for the individual model, and 43.02 mg/dL and 43.61 mg/dL for the general model, respectively. The independent components selected for the models were found to have spatial characteristics associated with variations in glucose concentrations in the orbital region and in the vessels on both sides of the nose.