Learning General Brain Network Representations of Different Brain Disorders Using Invariant Subgraph GNN

Distribution shifts across data from various brain disorders pose significant challenges for diagnosis. Establishing general feature representations that can handle these distribution shifts is crucial for accurately diagnosing these conditions. However, this area remains largely unexplored. This work propose an Invariant Subgraph GNN (IS-GNN) to learn general brain network representations for classifying various brain disorders in resting-state fMRI. This model employs an invariant subgraph learning mechanism to capture invariant brain graphs and handle distribution shifts. Moreover, we have developed an adaptive structure perception module to improve the detection of invariant subgraph features in brain networks by assessing the importance of nodes within the brain graph. To further refine the model, we propose a self-supervised loss for invariant subgraph learning, ensuring the generation of invariant brain network representations. Pretrained on data from 1,943 subjects across three public datasets corresponding to Autism Spectrum Disorder, Attention Deficit Hyperactivity Disorder, and Parkinson's Disease, the fine-tuning experiments of our proposed method demonstrate that the model achieves the state-of-the-art classification performance on not only the three datasets but also on an external Alzheimer's Disease dataset across.