Semi-supervised multimodal classification of Alzheimer's disease

One challenge in identification of Alzheimer's disease (AD) is that the number of AD patients and healthy controls (HCs) is generally very small, thus difficult to train a powerful AD classifier. On the other hand, besides AD and HC subjects, we often have MR brain images available from other related subjects such as those with mild cognitive impairment (MCI), a prodromal stage of AD, or possibly the unrelated subjects whose cognitive statuses may be not known. These images may be helpful for building a powerful AD classifier, although their cognitive status may not belong to AD or HC. Accordingly, in this paper, we investigate the potential of using MCI subjects to aid classification of AD from HC subjects via multimodal imaging data and CSF biomarkers. In particular, a multimodal Laplacian Regularized Least Squares (mLapRLS) method, based on semi-supervised learning, is proposed for achieving this purpose. In the objective function of mLapRLS, there are two terms: a term involving only AD and HC subjects for supervised learning, and another term involving all AD, HC, and MCI subjects for unsupervised estimation of intrinsic geometric structure of the data. Experimental results show that our proposed method can significantly improve AD classification, with aid from MCI subjects.