Graph neural network-based modeling of functional brain connectivity in autism spectrum disorder using the ABIDE-I and II datasets.
TL;DR
Three representative graph neural network architectures for ASD classification based on resting-state functional MRI reveal that distinct GNN architectures converge on common neurobiological signatures of ASD, highlighting their potential for reliable brain network modeling and biomarker discovery in psychiatry research.
Abstract
Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by atypical large-scale brain network organization. Using the comprehensive Autism Brain Imaging Data Exchange (ABIDE-I and ABIDE-II; n = 2013 participants), this study systematically evaluated three representative graph neural network (GNN) architectures for ASD classification based on resting-state functional MRI. The Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Self-Attention Graph Pooling (SAGPool) models were applied to graph representations of functional connectivity among 200 brain regions, following standardized preprocessing and ComBat harmonization. Across 100 repetitions of 10-fold stratified cross-validation, all three models achieved consistent and reproducible accuracy (∼61% balanced accuracy), establishing a robust performance benchmark for connectome-based ASD classification. Permutation-based feature importance analysis revealed convergent brain regions across models, including the bilateral thalamus, right superior temporal gyrus, right middle occipital gyrus, and bilateral precuneus, which are involved in sensory integration, social cognition, and default mode network functioning. These findings indicate that distinct GNN architectures converge on common neurobiological signatures of ASD, highlighting their potential for reliable brain network modeling and biomarker discovery in psychiatry research.