TRACKING ALZHEIMER’S DISEASE PROGRESSION BY NON-LINEAR DIMENSION REDUCTION OF BRAIN MRI FEATURES

Abstract

In addition to the necessary high diagnostic accuracy, an overall global index summarizing multiple complicated neuroimage based features of Alzheimer's disease (AD) should also be practically intuitive and logically explainable. In this research, we propose a new global index, which is derived from non-linear dimension reduction of brain MRI features, to track AD progression. In this research, we apply locally linear embedding (LLE) to a dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, which includes total of 346 volumes of brain regions and the cortical thickness of 562 subjects. Each subject has 2 to 8 MRI scans over time and about 20% of them have progressed to the next level of dementia. Among the original features, 59 most important features to the diagnosis are selected after prescreening. Based on the baseline data of 177 Cognitively Normal (CN) and 110 AD subjects, LLE can reduce the feature dimension to two and the probability of belonging to AD category can be assigned to each subject by a classifier. Using this baseline template, any subject's progression path can be depicted in the two-dimensional LLE feature space with his/her probability estimated from the nearest neighbors of baseline LLE template. Figure 1 shows the baseline template, which is constructed by two LLE coordinates and the probability obtained by Support Vector Machines (SVM) with Gaussian kernel with a sensitivity of 0.836 and a specificity of 0.955. Figure 2 illustrates a patient's disease progression path from Mild Cognitive Impairment (MCI) to AD. The probability of this patient at each visit is estimated by a linear combination of 16 nearest neighbors over baseline LLE template.