Noninvasive determination of blood glucose level based on short-wave NIR spectroscopy and simple liner regression

Recently people are growing conscious about their health and require minimum or non-invasive diagnostics. Non-invasive diagnosis of blood glucose level has long been expected to be a practical application, however unrealized. It is very difficult to build a universal prediction model from noninvasively measured NIR spectra not only for large number of people but also for different dates even for a specific person owing to the structure of blood vessel pattern or daily physical condition. The purpose of this paper is to propose a simple idea for calibrating daily and tailor-made prediction models for blood glucose level.
The glucose tolerance test was designed for determination of glycemic index (GI) for different kinds of vegetable foods with the aseptic packed Satou Rice based on the protocol of Japan association for study of GI. Both of conventional invasive blood glucose measurements with a SMBG meter and noninvasive NIR measurements were done at almost the same time seven times in two hours sequence including fasting condition (0: fasting, 15, 30, 45, 60, 90, and 120 minutes after first bite). A total of 34 healthy Japanese female students underwent repeatedly at least five times in different dates and totally 385 tolerance tests were done. As the noninvasive method, short-wave NIR spectra (700-1050 nm) were measured from hand palm by an interactance probe. Five datasets (0, 15, 30, 45, and 120 min) were selected for building a calibration model and the remained two were used for testing the model. The second order derivative spectra were calculated and the differences of the intensities from that of the fasting condition were used as input variables. Consequently we discovered the wavelength showing the highest correlation to the corresponding blood glucose levels measured by a SMBG meter fluctuates for every tolerance test. Therefore, the calibration model is necessarily a daily and tailor-made model. The calibration model was built using weighted sums of the input variables with the correlation coefficients as explanatory variables.
The standard error of prediction of blood glucose levels was 21.9 mg/dL for the data obtained in 60 and 90 minutes after eating (total number of data was 770) and the 95% confidence interval of standard error was ranged from -42.1 to 45.2 mg/dL. The gap between the actual and predicted GI values was practically small (ranged from 0.4 to 5.9) and enough to detect a significant difference of the vegetables. Our approach that can reduce pains in collecting blood may be effective for continuous monitoring of blood glucose level during limited hours as well as determination of GI.