Stress-Testing Hierarchical Brain Graph Decoding on Multi-Site rs-fMRI

Graph-based modeling of resting-state functional MRI (rs-fMRI) provides a promising framework for learning functional brain organization for psychiatric classification, but its generalization beyond binary diagnostic settings remains underexplored. We evaluate an age-conditioned hierarchical brain graph model on five-way psychiatric decoding using the multi-site SRPBS_OPEN cohort ((n=942); healthy controls (HC), autism spectrum disorder (ASD), major depressive disorder (MDD), schizophrenia (SCZ), and bipolar disorder (BD). Two rs-fMRI representations are investigated: ROI BOLD time series and dynamic amplitude of low-frequency fluctuation (dALFF) combined with static functional connectivity (FC). Under stratified five-fold cross-validation, BOLD achieved a balanced accuracy of (0.22+/- 0.02) and macro-AUC of (0.61+/-0.02), indicating modest discriminative ranking despite imbalanced class predictions. In contrast, dALFF+FC converged toward the HC majority class, with a balanced accuracy of (0.20) and macro-AUC of (0.50). These results highlight the challenges of transferring brain graph models from comparatively homogeneous binary benchmarks to heterogeneous multi-disorder settings and demonstrate that model behavior is sensitive to the input representation. The findings motivate extending hierarchical brain graph models with disorder-sensitive mechanisms that explicitly capture altered functional brain organization, potentially enabling more robust representation learning and generalization across psychiatric conditions.