Glucose concentrations in the interstitial fluid were measured based on optical spectroscopy. Prediction of glucose concentrations was made using partial least squares regression and accuracy improvement was achieved by data preprocessing as well as by selecting an optimal wavelength region. For this purpose, artificial interstitial fluid samples were prepared where their glucose levels varied between 0 and 10 g/dl. Infrared spectral regions where glucose absorption lies were investigated. A region of 1000 - 1500 $cm^{-1}$ produced the best accuracy among the regions of 1000 - 1500 $cm^{-1}$, 4000 - 4545 $cm^{-1}$1 and 5500 - 6500 $cm^{-1}$. Further accuracy improvement in 1000 - 1500 $cm^{-1}$ was achieved by selecting specific wavelength bands based on a loading vector analysis method. For the samples whose glucose concentrations ranged between 0 and 0.5 g/dl, SEP= 0.0266 g/dl and R =0.9863 were achieved with 1000 - 1500 $cm^{-1}$. However, the loading vector optimized band of 1002 - 1095 $cm^{-1}$ reduced the prediction error up to 47 % (SEP =0.0125 g/dl and R=0.9970).