Effective biomarkers play important roles for accurate diagnosis of Alzheimer’s Disease (AD), including its intermediate stage (i.e. mild cognitive impairment, MCI). In this paper, a new feature selection scheme was proposed to improve the identification AD and MCI from healthy controls (HC) by a support vector machine (SVM) based-classifier with recursive feature addition. Our method can find the significant features automatically, and the experiments in this work demonstrates that our scheme can achieve better classification performance based on a dataset with 103 subjects where three biomarkers, i.e., structural MR imaging (MRI), functional imaging PET, and cerebrospinal fluid(CSF), had been used. Our proposed method demonstrated its effectiveness in identifying AD from HC with an accuracy of 95.0%, while only 89.3% for the classifier without the step of feature selection. In addition, some features selected in this work had shown strong relation with AD by other previous studies, which can provide the support for the significance of our results.