Early Parkinson’s Disease Prediction Using rS-fMRI Functional Connectivity and Autoencoder Graph Convolutional Network

Early identification of prodromal Parkinson’s disease (PD) is critical, as interventions at this stage can significantly alter its course. We propose a deep learning framework that combines resting-state functional MRI (rs-fMRI) data and a Graph Convolutional Network (GCN) to classify individuals with PD, prodromal PD, and healthy controls. Our dataset consisted of 908 participants from two publicly available sources (Parkinson’s Progression Markers Initiative and SRPBS1600), including 288 with PD, 103 with prodromal (expanded to 309 via data augmentation), and 311 controls. Functional connectivity (FC) features were extracted using the Bootstrap Analysis of Stable Clusters (BASC) atlas by, applying different connectivity measures: Pearson’s, partial, Spearman’s, and tangent-space correlations. An autoencoder was then used to compress these features into a lower-dimensional space. Next, we constructed graph adjacency matrices using a novel neighborhood-based method and different distance metrics (Euclidean, spectral angle mapper, spectral information divergence, and radial basis function). This approach links each subject only to its most similar neighbors, yielding a sparse set of connections that preserves both local (nearest-neighbor) relationships and allows the GCN to capture global interactions across the entire dataset. Traditional classifiers (support vector machine, logistic regression, and random forest) demonstrated accuracies of 85.71%, 84.83%, and 80.77%, respectively. By contrast, the GCN trained with tangent-space correlation FC and using the radial basis function and Euclidean distance metric for adjacency matrix construction achieved higher accuracies of 90.66% and 90.64%. These findings underscore the effectiveness of tangent-space FC and GCNs for the early detection of PD.