Learning Dynamic Brain Network Representation Based on Graph Mamba Architecture

Learning graph representations of brain networks aims to facilitate a deeper understanding of the brain’s functional organizational patterns by developing data-driven algorithms on graph structures, thereby advancing neuroscience. Recently, deep learning models for brain networks based on graph neural networks (GNN) have emerged and become predominant in several tasks of brain functional analysis. Despite the significant successes of graph representation learning in brain networks, there remains a gap in dynamically modeling brain networks as they change over time. Inspired by state space modeling (SSM) in large language models, particularly Mamba, we introduce brain graph Mamba (Brain-GM) for dynamic brain graph representation learning. Specifically, we first validated the brain decoding performance of the Brain-GM model on the HCP dataset. Subsequently, we achieved leading disease diagnostic outcomes on unimodal brain disease datasets (ABIDE and ADHD), and for multi-modal neuroimaging data, we further optimized classification performance through similarity network fusion (SNF). Brain-GM marks the first integration of the state space model into dynamic brain graph representation learning, where the model’s graph learning readout layer and state-space features also support interpretability, enhancing our understanding of brain mechanisms related to different task states or diseases. The experimental code has been open-sourced at https://github.com/tokamak-wang/brain-gm.