Research on migraine classification model based on hypergraph neural network

Migraine is a common chronic neurological disorder that lacks objective imaging biomarkers, while resting-state functional magnetic resonance imaging (rs-fMRI) can be used to extract potential biomarkers. Recently, graph neural networks (GNNs) have gained significant popularity in the classification of brain disorders because of their powerful ability to model brains. Hypergraph neural networks (HGNNs), a branch of GNNs, are particularly effective in capturing high-order neighborhood information. In this paper, we proposed a hypergraph neural network model, incorporating hypergraph dual attention mechanism and hypergraph pooling strategy (APHGNN), for migraine classification derived from the preprocessed rs-fMRI data. First, we constructed hypergraphs from functional connectivity matrices based on the preprocessed rs-fMRI data. Then, we designed three network layers: in the hypergraph dual attention layer, we introduced attention mechanism in both the hyperedge feature aggregation phase and the node feature aggregation phase, making full use of both node and hyperedge information to update node features; in the hypergraph pooling layer, we employed a node selection-based pooling strategy to score and filter nodes, retaining key node features; in the readout layer, we calculated the average and maximum values of the key node features, concatenated and aggregated them, and used the resulting vectors for classification. The experimental results demonstrate that our model outperforms other baseline methods in classification performance and exhibits good generalization. Additionally, the key brain regions extracted through the hypergraph pooling strategy can serve as potential biomarkers for migraine, providing valuable insights for migraine diagnosis.