Multimodal Population Graph Networks with Self-Attention Mechanisms for Autism Spectrum Disorder Identification

Autism Spectrum Disorder (ASD) identification is challenged by data heterogeneity across sites, unclear population-level inter-subject relationship modeling, and a lack of biologically meaningful brain-region analysis. To address these issues, this paper proposes GPGATNet, a cross-site ASD identification framework that introduces learnable mechanisms at three levels-individual representation alignment, population graph construction, and stable propagation learning-to achieve robust classification while enabling biomarker discovery. A three-layer self-attention unsupervised graph pooling module, SAGNet, is designed to adaptively extract brain-feature subgraphs via node-importance scoring and hierarchical Top-k selection. This strategy alleviates cross-site scale discrepancies and simultaneously provides interpretable estimates of regional contributions. In addition, a population-graph modeling scheme based on a "phenotypic skeleton + functional-connectivity edge-weight updating" paradigm is employed to enhance the robustness of inter-subject relationship modeling. Furthermore, a graph convolutional network, MH-GCAT, is constructed by integrating multi-branch convolutions with multi-head attention fusion to strengthen population-level discriminative capability. Experiments on the ABIDE dataset demonstrate that GPGATNet outperforms existing methods. Moreover, by combining attention-based localization with analyses of BOLD signal fluctuations, we observe functional abnormalities and activity instability in key regions such as the thalamus and hippocampus in ASD, providing neuroimaging evidence for understanding impairments in emotion regulation and social behavior.