A Graph Neural Network Approach to Parkinson’s Disease Detection with Multi-Modal Brain Imaging

Abstract

Parkinson’s Disease (PD) is a progressive neurodegenerative disorder characterized by the grad ual loss of dopamine-producing neurons in the substantia nigra, resulting in both motor symptoms such as tremors, rigidity, and bradykinesia, and non-motor symptoms including cognitive decline and sleep distur bances. Early detection of PD remains a significant challenge due to the absence of reliable biomarkers prior to the onset of motor symptoms. Consequently, diagnosis often occurs only after substantial neuronal loss, reducing the efficacy of treatment. According to the World Health Organization (WHO), over 8.5 million people worldwide were affected by PD in 2023, emphasizing its growing global burden and the urgent need for improved diagnostic methods. Existing neuroimaging-based PD detection studies commonly employ Convo lutional Neural Networks (CNNs) that analyze voxel-level information, which limits their ability to capture the brain’s complex networked organization. Since PD disrupts both structural and functional connectivity across brain regions, representing the brain as a graph provides a biologically meaningful way to model inter-regional relationships. In this study, we propose a Graph Neural Network (GNN) based multi-modal framework for PD detection using diffusion-weighted MRI (dwMRI) and resting-state fMRI (rsfMRI) data from 115 PD and 118 normal control (NC) participants. Structural and functional brain networks were constructed for each modality, and five node-level features including degree, degree mean, external edges, ex ternal edges mean, and internal edges were extracted from all nodes in each subject. In functional networks, positive and negative edge weights were separately modeled through two GNN pipelines, and their outputs were ensembled. The structural and functional GNNs achieved accuracies of 88% and 75%, respectively, while the multi-modal ensemble improved classification performance to 90%. These findings demonstrate the potential of GNN based brain network analysis for effectively capturing PD related connectivity alter ations and advancing neuroimaging based diagnostics.