Machine Learning of Brain Functional Network Characteristics for AD Classification

Machine learning (ML) approaches have been employed to explore the feasibility of brain functional network characteristics as biomarkers, and the performance of MLs still needed to be verified. In our study, the brain functional network characteristics were investigated for the performance evaluation on the ML classifications among the three groups of normal control (NC), mild cognitive impairment (MCI) and Alzheimer's disease (AD). A cohort of 345 subjects enrolled for resting-state functional magnetic resonance imaging (rs-fMRI) data acquisition was divided into the three groups of NC, MCI and AD. Five brain functional network characteristics including path length (Lp), clustering coefficient (Cp), local efficiency (El), global efficiency (Eg) and small-worldness (σ), and four clinical scales of MoCA, FAQ, CDR-SB and MMSE were used in the classification of five MLs of Bagging, XGBoost, AdaBoost, Decision Trees and Random Forest. With the characteristic optimization, the classification performance of MLs using functional network characteristics and joint network characteristics was compared by the metrics of accuracy (ACC), area under the curve (AUC), precision (PRE), recall (REC) and F1. The joint characteristics had obtained better performance with the ACC of 91.46% for AD vs NC, 85.90% for AD vs MCI, 90.60% for MCI vs NC, and 74.64% for multi-class classification respectively. It was concluded that the MLs with combined brain characteristics exhibited superior performance for AD classification. Furthermore, the optimized brain characteristics hold potential as valuable biomarkers for early AD diagnosis.