Classification of Alzheimer's disease and mild cognitive impairment: Machine learning applied to rs-fMRI brain graphs

A growing number of studies use resting state functional magnetic resonance imaging (rs-fMRI) to investigate functional alterations in Alzheimer's disease (AD) and Mild cognitive impairment (MCI). In this study, we evaluate the effectiveness of graph theory and machine learning in the diagnosis of AD and MCI at the subject level. Moreover, we explore the effect of different methods of graph construction on the classification accuracy. To this end, the rs-fMRI data from 32 AD, 76 MCI, and 46 healthy elderly subjects are used along with three different connectivity measures: Pearson correlation, bend correlation, and mutual information. Both weighted and binary graphs are created for each subject at different density ranges. Based on each, various graph measures have been extracted. Finally, sequential floating forward selection and support vector machine are utilized. Our results suggest that bend correlation is more effective in the classification of AD and MCI. Moreover, weighted graphs and lower density ranges increase accuracy. Finally, bend correlation along with the weighted graph and density threshold of 10% generate the highest accuracy (93%). This study concludes that rs-fMRI and graph theory may provide a non-invasive means for the diagnosis of AD and MCI.