Selective multivariate analysis of blood glucose with near infrared spectra

Abstract

Optical properties of whole bovine blood are examined under conditions of different glucose loadings. Partial least-squares (PLS) is used to compute calibration models for glucose from spectra collected over the combination spectral region (5000 - 4000 cm- 1) and first overtone - short wavelength spectral regions (9000 - 5400 cm-1). These models achieve a prediction accuracy of approximately 1mM. Calibration models built for specific glucose absorption regions perform better than models generated strictly from the short wavelength region in which light scattering effects dominate. Net analyte signal (NAS) analysis is employed to investigate the spectral information that forms the basis for the models. The NAS reveals the portion of the glucose spectrum that is orthogonal to the spectral variance induced by the blood matrix. To investigate the selectivity of the spectral measurements, the glucose NAS is compared to residual absorbance spectra formed after subtraction of the non-glucose variance (estimated by application of principal component analysis to a set of blood samples with endogenous glucose concentrations). A match between the NAS and the residual spectra reveals that direct information associated with absorption of light by the glucose molecule is present in the measured data. A similar comparison is made with the regression vector associated with the PLS model. A match between the NAS and regression vector confirms that the correlations encoded in the calibration model do, in fact, arise from glucose absorption information. The results obtained through this work demonstrate that NAS analysis is a valuable tool for use in investigating the selectivity of multivariate calibration models.