Comparison of Multivariate Regression Models Based on Water- and Carbohydrate-Related Spectral Regions in the Near-Infrared for Aqueous Solutions of Glucose

The predictive power of the two major water bands centered at 6900 cm(-1) and 5200 cm(-1) in the near-infrared (NIR) region was compared to carbohydrate-related spectral areas located in the first overtone (around 6000 cm(-1)) and combination (around 4500 cm(-1)) region using glucose in aqueous solutions as a model substance. For the purpose of optimal coverage of stronger as well as weaker absorbing NIR regions, cells with three different declared optical pathlengths were employed. The sample set consisted of multiple separately prepared batches in the range of 50-200 mmol/L. Moreover, the samples were divided into a calibration set for the construction of the partial least squares regression (PLS-R) models and a test set for the validation process with independent samples. The first overtone and combination region showed relative prediction errors between 0.4-1.6% with only one PLS-R factor required. On the other hand, the errors for the water bands were found between 1.6-8.3% and up to three PLS-R factors required. The best PLS-R models resulted from the cell with 1 mm optical pathlength. In general, the results suggested that the carbohydrate-related regions in the first overtone and combination region should be preferred over the regions of the two dominant water bands.