It is important to identify people at high risk of diabetes, such as correlating them from physical examination data, and to intervene in advance in health management and medical treatment. In this paper, we employ both bagging and boosting methods to predict the patient’s fasting blood glucose (FBG) based on physical examination data. A hierarchical feature selection method based on sequential backward selection (SBS) algorithm is presented to select an optimal feature subset. The results of extensive experiments on account of physical examination database suggest that the presented feature selection method has better performance.