Processing of near-infrared multi-wavelength and multi-point data for glucose concentration measurement

Abstract

It is necessary for diabetic patients to measure blood glucose levels with blood sampling several times a day. Thus, non-invasive blood glucose sensors have been required to avoid the patients’ psychological and physical stress and the risk of infection. However, for non-invasive techniques proposed so far such as near-infrared (NIR) spectroscopy, the glucose signal tends to be buried due to the complexity in human skin structure and individual difference. To solve these problems, we have developed a NIR measurement device using multiple wavelengths and measurment points, and a regression model based on a neural network. In this paper, the relationship between the NIR absorbance data of subjects and glucose concentrations aquired by blood sampling was learned by a neural network, and moreover the transfer learning was performed to correct the bias of the measured data due to individual differences. It is found that this method is effective for correcting the bias for each subject and improving the accuracy of the regression model.