Evaluation of Data Pretreatment and Model Building Methods for the Determination of Glucose from Near-Infrared Single-Beam Spectra

Abstract

Near-infrared single-beam spectra are used to build partial least-squares (PLS) calibration models for the determination of glucose in biological matrices. Two different data sets of the same sample constituents are used in this investigation. The glucose samples consist of an aqueous matrix of varied concentrations of bovine serum albumin (BSA) and triacetin. The BSA and triacetin are models for blood proteins and triglycerides, respectively. Due to the effects of intensity variation in the single-beam spectra, calibration models obtained with unprocessed spectra are not as good as those computed with the corresponding spectra in absorbance units. When this intensity variation is reduced through the use of multiplicative signal correction (MSC), a spectral normalization method, or a logarithmic transform, the resulting models are as good as or better than those obtained in the analysis of absorbance spectra. An attempt is made to model the nonlinear relationship between single-beam spectral intensities and glucose concentrations by use of step-wise quadratic PLS (QPLS) models. The QPLS models are found to perform better than linear PLS models in some cases (e.g., with MSC-corrected single-beam spectra). The effect of digital filtering on the calibration models computed with single-beam spectra is also studied. The results obtained with and without filtering are found to be similar in terms of model performance, but the models based on filtered single-beam intensities require fewer latent variables and perform more consistently as a group. A final test is performed to compare the robustness of calibration models computed with single-beam spectra to those based on absorbance spectra. When applied to spectra that lie outside the time span of the calibration data, the models based on single-beam spectra are still competitive with those computed with absorbance spectra.