Near-infrared diffuse reflectance spectroscopy is a promising approach to the non-invasive prediction of blood glucose levels. However, because the measured object is human body whose physiological structures are so complicated that it is very difficult to separate the information of glucose from the overlapped spectra using the traditional modeling method. A new regression method called orthogonal projections to latent structures (O-PLS), which integrated the orthogonal signal correction (OSC) preprocessing into the regular PLS modeling, was applied to the optimization and interpretation of the glucose near-infrared spectroscopy model. Applying O-PLS resulted in removal of non-correlated variation in spectra and reduced model complexity with preserved prediction ability, improved interpretative ability of variation in spectra.