TSCNet: transformer-simplicial convolutional network for functional brain network classification

Functional brain network modeling provides an important tool for analyzing disease-related abnormalities in functional connectivity. Conventional graph models, however, treat nodes and edges as their basic objects and mainly encode pairwise relationships between brain regions, making it difficult to explicitly represent higher-order topological units jointly formed by multiple regions. We therefore propose TSCNet, a functional brain network classification framework that integrates simplicial convolution with a Transformer. TSCNet first maps multi-site-harmonized functional connectivity matrices to hierarchical simplicial complexes containing 0-, 1-, and 2-simplices. Hodge-Laplacian-driven simplicial convolutions then extract local topological features separately at each order. These order-specific features are projected into a shared representation space and passed to a Transformer to model global dependencies across simplices and brain regions. We evaluated TSCNet on three public fMRI datasets, ABIDE, ADHD-200, and Haxby, using stratified 10-fold cross-validation repeated five times. The complete model achieved classification accuracies of 81.23% on ABIDE and 82.42% on ADHD-200. Removing either the SCNN or Transformer component reduced performance, indicating that local higher-order topology and global information integration provide complementary information. Parameter-sensitivity analyses and topological-pattern visualizations further characterized the model’s responses to simplex order and construction threshold. TSCNet provides a framework for jointly analyzing local higher-order topology and whole-brain dependencies, although its generalization to unseen scanners and the clinical specificity of the identified higher-order patterns require validation in independent cohorts.