FC–HAT: Hypergraph attention network for functional brain network classification

Abstract

Recently, functional brain network analysis via graph neural networks has achieved state-of-the-art results as it can directly extract information from irregular graphs without any approximation. However, current methods remain limited in exploring the high-order structural information of brain networks. To address this issue, we propose a hypergraph attention network for functional brain network classification (FC–HAT). First, we build a dynamic hypergraph generation phase and model each brain network as a hypergraph to preserve the high-order information. The pair-wise and community-wise similarities in functional brain networks are separately captured by k nearest neighbors and k-means. Theoretical analysis shows that the constructed hypergraph exhibits superior spectral properties. Then, we design a hypergraph attention aggregation phase to further extract information in hypergraphs. This includes node and hyperedge attention layers that can separately aggregate features among nodes and hyperedges. Finally, the two phases are jointly optimized in an end-to-end manner, which can dynamically update hypergraphs and node embeddings along with the training process. Experimental results on ABIDE-I and ADHD-200 demonstrate the effectiveness of FC–HAT in cerebral disease classification. Moreover, the abnormal connectivity patterns and brain regions identified by FC–HAT are found to be more likely to become biomarkers associated with cerebral diseases.