Gradient-Based Explainable Graph Attention Networks for Parkinson’s Disease Classification Using Feature-Informed Structural Brain Networks

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder caused by degeneration of dopaminergic neurons in the substantia nigra, leading to motor and non-motor impairments. Early diagnosis is important because timely intervention may improve symptom management and slow disease progression. However, subtle structural brain alterations make early detection difficult using conventional neuroimaging. Diffusion magnetic resonance imaging (dMRI) provides a non-invasive approach for detecting microstructural changes in white matter pathways. This study proposes a gradient-based explainable Graph Attention Network (GAT) for PD classification using feature-informed structural brain networks derived from dMRI. Weighted connectivity networks were constructed from 130 PD patients and 130 healthy controls from the Parkinson’s Progression Markers Initiative dataset. Each brain region was modeled as a node enriched with graph-theoretic features, while white matter connections formed weighted edges. The GAT learned discriminative disease patterns by adaptively aggregating information from connected neighboring regions. A gradient-based explainability module was integrated to identify the most influential brain regions contributing to classification. The proposed framework achieved 98% classification accuracy and identified clinically relevant regions, including the basal ganglia, insula, and motor cortex, consistent with known PD-related structural abnormalities.