Integrative analysis of multi-dimensional imaging genomics data for Alzheimer's disease prediction

In this paper, we explore the effects of integrating multi-dimensional imaging genomics data for Alzheimer's disease (AD) prediction using machine learning approaches. Precisely, we compare our three recent proposed feature selection methods [i.e., multiple kernel learning (MKL), high-order graph matching based feature selection (HGM-FS), sparse multimodal learning (SMML)] using four widely-used modalities [i.e., magnetic resonance imaging (MRI), positron emission tomography (PET), cerebrospinal fluid (CSF), and genetic modality single-nucleotide polymorphism (SNP)]. This study demonstrates the performance of each method using these modalities individually or integratively, and may be valuable to clinical tests in practice. Our experimental results suggest that for AD prediction, in general, (1) in terms of accuracy, PET is the best modality; (2) Even though the discriminant power of genetic SNP features is weak, adding this modality to other modalities does help improve the classification accuracy; (3) HGM-FS works best among the three feature selection methods; (4) Some of the selected features are shared by all the feature selection methods, which may have high correlation with the disease. Using all the modalities on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, the best accuracies, described as (mean +/- standard deviation)%, among the three methods are (76.2 +/- 11.3)% for AD vs. MCI, (94.8 +/- 7.3)% for AD vs. HC, (76.5 +/- 11.1)% for MCI vs. HC, and (71.0 +/- 8.4)% for AD vs. MCI vs. HC, respectively.