Interpretable Graph Neural Networks for Parkinsonian Brain Connectivity Analysis Using Resting-State fMRI
Main article
Abstract
Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder, affecting over 10 million people worldwide. Alterations in functional brain connectivity, particularly within the cortico-striato-thalamo-cortical circuits, represent a promising neuroimaging biomarker for PD diagnosis and progression monitoring. However, existing deep learning approaches for automated PD detection based on resting-state functional magnetic resonance imaging (rs-fMRI) brain networks suffer from limited interpretability, hindering their clinical adoption. This study proposes Interpretable Graph Neural Networks (IG-GNN), a novel framework that integrates graph contrastive pretraining, a variational graph encoder with attention mechanisms, and prototype-based subgraph interpretation to enable simultaneous high-accuracy classification and biologically meaningful explanations. We evaluate our framework on data from the Parkinson's Progression Markers Initiative (PPMI) dataset comprising 86 PD patients and 70 healthy controls. IG-GNN achieves an accuracy of 87.6%, sensitivity of 86.4%, specificity of 88.7%, and an AUC of 0.921, outperforming six state-of-the-art baseline methods. Interpretability analysis identifies abnormal connectivity in the putamen, caudate nucleus, supplementary motor area, and thalamus as the most discriminative features, consistent with established neuropathological evidence. Our results demonstrate that interpretable GNN architectures can provide clinically actionable diagnostic insights, bridging the gap between computational precision and neuroscientific understanding.
