BrainCHEF: Cross-Level Hypergraph Enhanced Fusion model for brain networks

Modeling the dynamic characteristics of functional brain networks is of great significance for uncovering the mechanisms of brain function. Although graph neural networks (GNNs) have achieved remarkable progress in the analysis of functional networks, they still face challenges in terms of data scarcity, insufficient supervision, and capturing high-order structural information. To address these challenges, we propose a Cross-Level Hypergraph-Enhanced Fusion Framework (BrainCHEF), which integrates the perspectives of hypergraphs and line graphs. The framework employs hypergraph attention networks to adaptively learn complex node dependencies and uses a self-supervised method with feature masking to enhance the capture of hyperedge interaction information in line graphs. By integrating global information through cross-level interaction mechanisms and combining persistent homology analysis to process fMRI signals for extracting dynamic structural information, BrainCHEF enhances the spatiotemporal characteristics of the model. The design of BrainCHEF not only improves the model's ability to capture the dynamic characteristics of brain networks but also enhances its generalization and interpretability. Extensive experiments on two real-world datasets, the Autism Brain Imaging Data Exchange (ABIDE) and Attention Deficit Hyperactivity Disorder (ADHD), demonstrate the superior performance of BrainCHEF, significantly outperforming existing state-of-the-art methods. Moreover, BrainCHEF is capable of identifying disease-related biomarkers that are consistent with previous research findings. Ablation studies further confirm the effectiveness and rationality of hypergraph modeling and self-supervised tasks, providing a new tool for the diagnosis and research of brain diseases. The relevant code has been made publicly available on GitHub: github.com/BrainCHEF.