selected for each combination model were used to construct the classification space using support vector machines. To combine MRI and PET, orthogonal partial least squares to latent structures is used as a multivariate analysis to discriminate between AD, early and late MCI (EMCI and LMCI) from cognitively normal (CN)s. In addition, this dissertation proposes a new effective mean indicator (EMI) method for distinguishing stages of AD from CN. EMI utilizes the mean of specific top-ranked measures, determined by incremental error analysis, to achieve optimal separation of AD and CN.viii For AD vs. CN, the two most discriminative volumetric variables (right hippocampus and left inferior lateral ventricle), when combined with MMSE scores, provided an average accuracy of 92.4% (sensitivity: 84.0%; specificity: 96.1%). MMSE scores were found to improve classification accuracy by 8.2% and 12% for aMCI vs. CN and naMCI vs. CN, respectively. Brain atrophy was almost evenly seen on both sides of the brain for AD subjects, which was different from right side dominance for aMCI and left side dominance for naMCI. Findings suggest that subcortical volume need not be normalized, whereas cortical thickness should be normalized either by intracranial volume or the mean thickness. Furthermore, MRI and PET had comparable predictive power in separating AD from CN. For the EMCI prediction, cortical thickness was found to be the best predictor, even better than using all features together. Validation with an external test set demonstrated that best of feature-selected models for the LMCI group was able to classify 83% of the LMCI subjects. The EMI-based method achieved an accuracy of 92.7% using only MRI features. The performance of the EMI-based method along with its simplicity suggests great potential for its use in clinical trials.