DB-STAGN: A Dual-Branch Spatio-Temporal Attention Graph Network for ADHD diagnosis from fMRI

Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder whose diagnosis is often limited by subjective clinical assessments. Deep learning analysis of functional magnetic resonance imaging (fMRI) data provides a promising avenue for identifying objective biomarkers. However, existing methods often struggle to effectively model the complex interplay among brain regions, particularly in simultaneously capturing long-range spatial dependencies and multi-scale temporal dynamics. To address these challenges, this study presents a Dual-Branch Spatio-Temporal Attention Graph Network (DB-STAGN) for ADHD diagnosis from fMRI-derived dynamic functional connectomes. The framework processes dynamic functional connectomes through two parallel branches: a spatial branch featuring a Topology-Guided Attention Module (TGAM), which integrates graph topological priors into a Transformer to learn global spatial patterns, and a temporal branch employing a Multi-Scale Temporal Convolutional Network (MS-TCN) to capture temporal dependencies at multiple resolutions. The learned features are then adaptively fused for classification. Experiments on three ADHD-200 sites demonstrate that DB-STAGN outperforms several representative baseline models under site-specific evaluation, with accuracies of 85.3% on NYU, 87.7% on NI, and 81.3% on KKI. In addition, visualization of the learned attention weights highlights discriminative brain regions such as the putamen, insula, and amygdala, which are broadly consistent with prior neurobiological findings in ADHD. These results suggest that DB-STAGN provides an effective and interpretable framework for automated ADHD diagnosis.