Task-Aware Functional Hypergraph Learning for Brain State Classification via Information Bottleneck

Functional connectivity networks (FCNs) are widely used in fMRI-based brain analysis. While most existing studies represent FCNs using graphs, traditional graph structures primarily focus on pairwise connections, overlooking higher-order relationships. Additionally, many methods construct graphs or hypergraphs independently of downstream tasks, which can result in suboptimal representations that fail to capture task-relevant structures. To address these limitations, we propose a novel approach that integrates task-specific information directly into the hypergraph construction process. Our method employs a learnable groupwise mask to construct a groupwise hypergraph structure across all subjects. To retain task-related brain regions and filter out irrelevant ones, we introduce an information bottleneck constraint to optimize our framework. Furthermore, to capture personalized information, we design a hypergraph multi-head attention mechanism that learns personalized hypergraph attention matrices. We apply our model to the ADNI-3 dataset and ABIDE dataset to classify brain states associated with Alzheimer's disease and autism. Our method outperforms competing approaches, achieving at least a $2.2\%$ improvement in accuracy.