HGMA-Net: A Hierarchical Graph-Mamba Encoder for Multi-Modal Brain Disorder Classification on Population Graphs.

Brain disorder (BD) classification based on population graph learning has shown promising performance by jointly modeling resting-state functional MRI (rs-fMRI) data and phenotypic information. Existing graph Transformer-based population graph encoders suffer from high computational cost on densely connected subject graphs and lack an efficient mechanism for global contextual modeling across distant subjects. To address these limitations, we propose HGMA-Net, a hierarchical graph-Mamba hybrid framework for multi-modal brain disorder classification. Specifically, the proposed HGMA module redesigns the shallow encoder stage using a dual-path architecture that combines local graph aggregation with global sequence modeling. A degree-guided node serialization strategy provides a topology-aware ordering of graph nodes, enabling bidirectional Mamba state space models to efficiently capture global contextual information with linear sequence complexity. An adaptive gating mechanism is further introduced to dynamically fuse global sequential features and local graph-structural features. To balance computational efficiency and representational capacity, HGMA is applied only at the first encoder layer, while deeper layers retain TransformerConv for local topology refinement. Experiments on the ABIDE and ADHD-200 datasets demonstrate that HGMA-Net achieves superior classification performance, obtaining AUC scores of 89.14% and 91.70%, respectively. HGMA-Net also retains efficient inference when global contextual modeling is introduced at the shallow encoder stage.