Evaluation of nonlinear model building strategies for the determination of glucose in biological matrices by near-infrared spectroscopy Original Research Article

Nonlinear model building techniques are applied to near-infrared spectra to predict glucose concentrations in samples containing an aqueous matrix of varied concentrations of bovine serum albumin (BSA) and triacetin. The triacetin is used to model triglycerides in human blood, and the BSA is used to model blood proteins. The non-linear model building techniques included in this study are quadratic partial least-squares regression (QPLS), stepwise QPLS, and PLS followed by artificial neural networks (PLS-ANN). The optimal models obtained for glucose provide standard errors of prediction of 0.53 mM, 0.54 mM, and 0.48 mM for the QPLS, stepwise QPLS and PLS-ANN models, respectively, over the clinically relevant concentration range of 1–20 mM. These results indicate significant improvement in prediction performance relative to that obtained with linear PLS models. This improvement is confirmed through the use of F-tests at the 95% confidence level. The significant quadratic terms included in the stepwise QPLS models also confirm that nonlinear information exists in the data set studied. This suggests that there is a need to develop suitable nonlinear model building strategies for noninvasive blood glucose determinations.