Sparse multimodal manifold-regularized transfer learning for MCI conversion prediction

Effective prediction of conversion of mild cognitive impairment (MCI) to Alzheimer's disease (AD) is important for early diagnosis of AD, as well as for evaluating AD risk pre-symptomatically. Different from most traditional methods for MCI conversion prediction, in this paper, we propose a novel sparse multimodal manifold-regularized transfer learning classification (SM2TLC) method, which can simultaneously use other related classification tasks (e.g., AD vs. normal controls (NC) classification) and also the unlabeled data for improving the MCI conversion prediction. Our proposed method includes two key components: (1) a criterion based on the maximum mean discrepancy (MMD) for eliminating the negative effect related to the distribution differences between the auxiliary (i.e., AD/NC) and the target (i.e., MCI converters/MCI non-converters) domains, and (2) a sparse semisupervised manifold-regularized least squares classification method for utilization of unlabeled data. Experimental results on the Alzheimer's Disease Neuroimaging Initiative (ADNI) database show that the proposed method can significantly improve the classification performance between MCI converters and MCI non-converters, compared with the state-of-the-art methods.