We made spectral measurements on a diverse group of 19 non-diabetic subjects. Measurements used diffuse reflectance, with a hemispherical fluorite lens and index matching fluid, from the underside of the left forearm. Spectra were collected from 10,000 wave numbers (1.0 microns) to 4,000 wave numbers (2.5 microns) for about three minutes each. Spectra and measurements, with a blood sample, of blood glucose concentration, were taken in fasting, elevated (after a carbohydrate snack) and post-snack conditions. We took three fasting spectra, six elevated spectra, and three post snack spectra. There were monotonic variations in the measured signal as a function of sample number. The multivariate analysis captured about 99 percent of the spectral variance in two principal components (PCs). The first component was a spectrum to spectrum variation in overall level. The second was a variable tilt to the spectra. Those PCs, which correlated best with glucose, captured less than 0.01 percent of the spectral variance. Our calibration model was poor. The error in prediction based on cross validation was on the order of 35 mg/dL. The standard deviation of the reference measurements (which were not normally distributed) was 31.6 mg/dL. One can conclude that our calibration model had no predictive value. 10 Pages
PDF
download
878987797
028-85220240
