The study of brain networks by resting-state functional magnetic resonance imaging
(rs-fMRI) is a promising method for identifying patients with dementia from healthy
controls (HC). Using graph theory, different aspects of the brain network can be efficiently
characterized by calculating measures of integration and segregation. In this study,
we combined a graph theoretical approach with advanced machine learning methods to
study the brain network in 89 patients with mild cognitive impairment (MCI), 34 patients
with Alzheimer's disease (AD), and 45 age-matched HC. The rs-fMRI connectivity matrix
was constructed using a brain parcellation based on a 264 putative functional areas.
Using the optimal features extracted from the graph measures, we were able to accurately
classify three groups (i.e., HC, MCI, and AD) with accuracy of 88.4 %. We also investigated
performance of our proposed method for a binary classification of a group (e.g., MCI)
from two other groups (e.g., HC and AD). The classification accuracies for identifying
HC from AD and MCI, AD from HC and MCI, and MCI from HC and AD, were 87.3, 97.5, and
72.0 %, respectively. In addition, results based on the parcellation of 264 regions
were compared to that of the automated anatomical labeling atlas (AAL), consisted
of 90 regions. The accuracy of classification of three groups using AAL was degraded
to 83.2 %. Our results show that combining the graph measures with the machine learning
approach, on the basis of the rs-fMRI connectivity analysis, may assist in diagnosis
of AD and MCI.