Ensemble Universum SVM Learning for Multimodal Classification of Alzheimer's Disease

Recently, machine learning methods (e.g., support vector machine (SVM)) have received increasing attentions in neuroimaging-based Alzheimer's disease (AD) classification studies. For classifying AD patients from normal controls (NC), standard SVM trains a classification model from only AD and NC subjects. However, in practice besides AD and NC subjects, there may also exist other subjects such as those with mild cognitive impairment (MCI). In this paper, we investigate the potential of using MCI subjects to aid the identification of AD from NC subjects. Specifically, we propose to use the universum support vector machine (U-SVM) learning by treating MCI subjects as the universum examples that do not belong to either of the classes (i.e., AD and NC) of interest. The idea of U-SVM learning is to separate AD from NC subjects through large margin hyperplane with the universum MCI subjects laying inside the margin borders, which is in accordance with our domain knowledge that MCI is a prodromal stage of AD with cognitive status between NC and AD. Furthermore, we propose ensemble universum SVM learning for multimodal classification by training an individual U-SVM classifier for each modality. Experimental results on the Alzheimer's Disease Neuroimaging Initiative (ADNI) database demonstrate the efficacy of our proposed method.