In order to select optimal spectral width for biomedical components analysis by near-infrared (NIR) spectroscopy, starting with the characteristic wavelengths of the samples, a modeling procedure based on spectroscopy principles is presented. 348 human serum samples are collected. Cholesterol and triglyceride in human sera are analyzed as an example. According to spectroscopy principles and absorbance of 2 mm-thick sera, first overtone region is selected for analysis. Models with different spectral width are compared. For cholesterol and triglyceride models, optimal spectral width is 70nm and 100nm respectively. Root mean square error of prediction (RMSEP) and mean percent error of prediction (MPEP) are 017 mmol/L, 3.0% and 0.14mmol/L, 10.3%, respectively. The performance is comparable with the best works of other groups while the spectral width is narrower.