This study explores the combined impact of cognitive and brain reserves with resting-state functional magnetic resonance imaging (rs-fMRI) based functional connectivity on Parkinson's Disease (PD) classification. A machine learning pipeline is presented aimed at discriminating between 52 healthy controls and 43 subjects with PD using a support vector machine (SVM) classifier trained using a 5-fold cross-validation approach. Augmenting our approach with an eXplainable artificial intelligence (XAI) tool, specifically the SHapley Additive exPlanation (SHAP) method for feature ranking, we explained the underlying mechanisms guiding the model decision. The results showed an average accuracy of 94.74% using the top 20 features with the highest SHAP importance score. Specific connections, such as those governing visual central and dorsal attention, emerged as key discriminative features, significantly impacting on the model's ability to classify PD subjects.
