Early classification of functional connectomes in Parkinson’s disease: a comparison of machine learning classifiers using multi-scale topological features

Parkinson’s disease (PD) affects brain networks across multiple anatomical scales, but it is unclear how correctly machine learning (ML) would classify the (altered) topology of functional connectomes in single PD cases at an early clinical stage. Leveraging network graph theory (NGT) for multi-scale topological feature extraction, here we determined (i) which features (across scales, lobes and types) are relatively more important, and (ii) which current state-of-the-art ML classifiers provide better performances, when discriminating newly diagnosed, drug-naive PD patients vs. healthy control (HC) subjects in a single-center MRI study. Resting-state functional MRI was consecutively performed in 112 drug-naïve PD patients and 17 HC subjects. Following standard (automated) preprocessing, the Brainnetome atlas (210/36 cortical/subcortical areas) was applied to reconstruct individual functional connectomes. NGT features were extracted at global, lobar/cortical and lobar/subcortical scales. Feature importance was assessed as the information gain (IG) criterion. The synthetic minority oversampling technique (SMOTE) was used to augment HC datasets. Nine different ML classifiers were trained on discriminating PD vs. HC (balanced) classes and validated via repeated stratified cross-validation. Among global NGT features, characteristic path length (IG = 20.3%) was more important than global (IG ~ 14%) and average local (IG ~ 14%) efficiency. Basal ganglia (IG = 18.5%), parietal cortex (IG = 17.7%) and thalamus (IG = 15.6%) cumulatively contributed the most informative NGT features. The “extra-tree classifier” achieved the best performances (F1-score = 95.8 ± 3.8%; AUCROC = 99.3 ± 1.8%). Both global and local NGT features are important for the topological classification of functional connectomes in drug-naïve PD patients. Combining multi-scale functional connectomics with ML may contribute to the design of health information technologies (for decision-making) and for the definition of early, accurate and interpretable, non-invasive neuroimaging PD biomarkers.