Multi-hop spatio-temporal graph convolutional networks for brain disorder diagnosis and prognosis

Published in Pattern Recognition (SCI Q1, IF 9.1), 2026

Abstract

Existing GNN attention methods for brain network analysis only aggregate one-hop neighbors, ignoring long-range indirect connections in sparse brain networks, while multi-site fMRI heterogeneity limits cross-site generalization. We propose MSTGCN, which extends spatial attention to multi-hop connections via geometric-distribution-weighted attention diffusion, employs a Transformer encoder for dynamic functional connectivity temporal modeling, and uses RevCon regularization to suppress site-specific biases. The model achieves state-of-the-art performance on multi-site fMRI datasets for brain disorder diagnosis and surgical prognosis.

Recommended citation: Liu, H.†, Zhang, J.†, Fang, Z.†, Jiang, X.*, Huang, W.*, Li, R.*. (Accepted). Multi-hop spatio-temporal graph convolutional networks for brain disorder diagnosis and prognosis. Pattern Recognition.
Download Paper