Multimodal Hypergraph Attention-Based Brain Network Analysis for Brain Diseases Diagnosis

Brain networks, represented as graphs modeling inter-regional interactions, are crucial for understanding abnormal connectivity patterns associated with neurological disorders. Existing methods typically leverage Graph Convolutional Networks (GCNs) to extract shallow topological features by aggregating local neighborhood information. However, neuroscience research highlights the brain's modular organization, where tightly connected regions within modules support complex cognitive functions. To capture such high-order modular interactions, we propose a novel hypergraph attention network framework. We first construct multimodal brain networks using fMRI and DTI data. Then, based on the interplay between functional and structural connectivity, we dynamically determine the number of clusters to build a hypergraph reflecting brain modularity. Finally, we integrate GCNs and attention mechanisms on this hypergraph to learn discriminative node representations. Extensive experiments on an epilepsy dataset demonstrate the effectiveness of our method in both classification performance and biomarker identification.