Maximum mutual information for feature extraction from graph-structured data: Application to Alzheimer's disease classification

A brain network can be constructed from various imaging modalities such as magnetic resonance imaging (MRI), representing the functional or structural connectivity between brain regions. The challenge of brain network analysis is efficient dimensionality reduction while retaining feature interpretability. We propose a new method to extract features from graph-structured data based on maximum mutual information (MMI-GSD). First, we develop a novel equation for the feature extraction from GSD and evaluate the interpretability of the features. We establish a framework to optimize the extracted features using the MMI. We conduct experiments on synthetic networks to validate the effectiveness of the proposed MMI-GSD. Next, we conduct experiments on 119 cognitively normal (CN), 105 mild cognitive impairment (MCI), and 36 Alzheimer's disease (AD) individuals from the Alzheimer's Disease Neuroimaging Initiative. The classification performance of the proposed method is significantly better than using traditional network metrics and existing feature extraction methods. In the clinical interpretation, we discover discriminative brain regions showing significant differences between the MCI and AD groups and identify significant abnormal connections concentrated in the left hemisphere.