Selective calibration models are generated for glucose over the 1-20 nM concentration range by use of partial least-squares regression analysis of near-infrared spectra from 5000 to 4000 cm-1. Two spectral data sets are used to simulate triglyceride and protein variations in clinical samples. Triacetin is used in one data set to simulate variations in triglyceride levels, and bovine serum albumin (BSA) is used in the second data set to simulate variations in blood protein levels. Although these matrix components possess strong absorption bands that overlap and overshadow the absorption bands of glucose, successful calibration models can be generated with no evidence of prediction bias caused by the different levels of the matrix components. Furthermore, the benefits of using digital Fourier filtering as a preprocessing step are evaluated in terms of calibration performance. The resulting calibration models provide standard errors of prediction of 0.5 and 0.2 mM in triacetin and BSA matrices, respectively. Accurate glucose predictions are demonstrated from spectra that correspond to protein concentrations not present in the calibration data set. Lastly, digital Fourier filtering alone is shown to have only limited ability to isolate glucose signals from those of BSA and triacetin due to similarities in the widths of the absorption bands of the three species.