Multivariate determination of glucose concentrations from optimally filtered frequency-warped NIR spectra of human blood serum

Abstract

Glucose concentrations over the 39 - 160 mg range have been determined from 357 NIR (near-infrared) spectra of human blood serum in the spectral region from 6766 to 4003 . A frequency-warping procedure was applied to the NIR data to compress 511 spectral components into 102 in the 6766 - 4003 spectral region. Before the data compression process was carried out, the NIR spectrum of deionized water was subtracted from each of the blood serum spectra to remove the intrinsic high background absorption due to the water. PLS (partial least-squares) regression was coupled with time-domain digital Butterworth bandpass filtering in an optimization procedure. The optimization procedure was carried out over a range of centre frequencies and bandwidths for first- (two-pole), second- (four-pole) and third- (six-pole) order bandpass filters, and over a range of PLS factors. The optimal PLS model and filter parameters were determined from a sequence of three-dimensional performance response maps for different numbers of PLS factors and filter orders. As a basis for comparison, the same optimization process was carried out for a Gaussian filter design approach (i.e., Fourier filtering). Using the optimally filtered frequency-warped NIR spectral data, an SEP (standard error of prediction ) of 13.2 mg was achieved for the test (monitoring) data using 14 PLS factors and a simple first-order (two-pole) digital Butterworth bandpass filter. Keywords: diabetes mellitus, glucose, human blood serum, non-invasive monitoring, near-infrared spectroscopy, partial least squares, optimal filtering, frequency warping