Mining Latent Connectivity Patterns in Parkinsonian Brain Networks through Variational Graph Representation Learning
Main article
Abstract
Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder globally, yet its diagnosis continues to rely on subjective clinical assessment, leading to delays and inconsistencies. Resting-state functional MRI (rs-fMRI) encodes rich information about brain connectivity alterations in PD, but extracting discriminative and interpretable patterns from high-dimensional brain graph data remains a substantial data mining challenge. This paper proposes a novel framework, LatentBrainNet, that mines latent connectivity patterns in Parkinsonian brain networks through variational graph representation learning. The framework integrates three stages: (i) multi-site graph contrastive pre-training using a Graph Convolutional Network (GCN) encoder to learn transferable brain graph representations; (ii) a Variational Graph Auto-Encoder (VGAE) that maps brain connectivity matrices into a low-dimensional probabilistic latent space, enabling uncertainty-aware feature extraction; and (iii) a prototype-based classifier that produces both discriminative predictions and subgraph-level explanations aligned with neurobiological knowledge. Evaluated on the PPMI dataset and an in-house multi-site fMRI cohort comprising 177 subjects, LatentBrainNet achieves 91.3% classification accuracy, 90.7% sensitivity, and an AUC of 0.952, surpassing competitive baselines by substantial margins. SHAP and subgraph explanation analysis consistently identify the basal ganglia–thalamus circuit, prefrontal–motor connections, and cerebellar networks as primary discriminative subgraphs, consistent with established PD neuropathology. The proposed data-driven framework advances automatic biomarker discovery and offers clinically interpretable decision support for PD diagnosis
