Multiscale integration of pairwise and higher-order fMRI-based functional connectivity in hypergraph neural networks enhances autism spectrum disorder classification

Conventional graph-based analyses of functional brain networks capture only pairwise interactions, overlooking higher-order polyadic dependencies that are increasingly recognized as fundamental to brain computation. Here we apply hypergraph neural networks (HGNNs) to resting-state fMRI data from the ABIDE dataset (408 ASD, 476 typically developing subjects; AAL atlas, 116 ROIs) and demonstrate that the synergistic integration of higher-order hypergraph structure with pairwise functional connectivity features is essential for accurate autism spectrum disorder (ASD) classification. Hypergraph representations are constructed via sparse group lasso, and two UniGNN variants (UniGCN and UniGIN) are benchmarked against SVM and standard GCN baselines. UniGIN with pairwise connectivity profiles as node features achieves the best performance (F1 = 0.77, AUC = 0.80, balanced accuracy = 0.76), substantially outperforming the same architecture with time-series-derived features (F1 = 0.64, AUC = 0.74) and all pairwise-only models. This result establishes that neither hypergraph topology alone nor pairwise connectivity alone suffices: the hypergraph defines meso-scale modular structure while pairwise features encode fine-grained dyadic interactions, and both scales are needed for optimal discrimination. Consensus network analysis reveals that group-level differences are structured and systematic at the hypergraph level but region-specific at the pairwise level. Contrastive comparison of consensus hypergraphs shows that 9 of 10 hyperedges are topologically identical between ASD and TD groups, sharing a conserved triple-network (DMN–CEN–SN) backbone. The primary divergence involves the cerebellum: ASD exhibits a decoupling of visual–cerebellar coordination, the emergence of an isolated cerebellar module with expanded Vermis recruitment, and an aberrant fronto-cerebellar hyperedge absent in controls. These findings provide a multi-scale framework for understanding ASD-related brain network alterations and establish a methodological principle that advancing beyond pairwise interactions requires integrating, not replacing, dyadic analysis within higher-order topological representations.