Classification of autism spectrum disorder using a directional graph attention network on brain effective connectivity.

Recent research into the neural mechanisms underlying autism spectrum increasingly relies on brain network analysis; however, conventional models remain limited in their ability to capture directed causal interactions between brain regions. To address this limitation, we propose a directional graph attention network (DGAT) as a proof-of-concept framework for autism classification using directed effective connectivity. DGAT takes Granger causality matrices as input and employs a dual-branch attention architecture to model incoming and outgoing information flows separately, then adaptively fuses the resulting bidirectional embeddings through learnable weights to more explicitly characterize driver-response relationships between regions. In addition, the model incorporates multiscale node descriptors, including temporal statistics, graph-theoretic centrality measures, and global graph metrics, to enhance representational capacity. Under nested cross-validation, DGAT achieved competitive performance on key metrics (accuracy: 71.99%, AUC: 75.15%, specificity: 73.07%) and produced more favorable results than support vector machines, random forests, graph convolutional networks, and GAT based on undirected functional connectivity. These findings suggest that DGAT may serve as a promising exploratory framework and offer a novel perspective for disease classification based on directed brain networks.