MD-DGNN: A Metadata-Driven Dual Graph Neural Network based on Brain Functional Connectivity for ASD Classification

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder that requires early and efficient diagnosis. Functional magnetic resonance imaging (fMRI), with high temporal and spatial resolution, is a powerful tool for investigating brain functional connectivity. In fMRI data-based ASD classification, the dual graph method integrates individual brain functional connectivity with metadata-based inter-subject relationships, achieving high diagnostic accuracy. However, relying on the predefined brain graphs with fixed structures in the brain graph learning stage fails to adaptively optimize graph topology for downstream tasks. Furthermore, existing methods perform early fusion of metadata to construct a single population graph, discarding critical metadata specific information and causing unclear inter-subject relationships. Accordingly, we propose a metadata-driven dual graph neural network (MD-DGNN) to alleviate these problems, which consists of an edge-adaptive GNN and a metadata-aware GNN to process the brain graphs and population graphs, respectively. In edge-adaptive GNN, we design a dynamic edge sampling module to adaptively learn the task-optimal graph topologies. In metadata-aware GNN, we construct heterogeneous population graphs based on different types of metadata and implement a hierarchical cross-fusion strategy for integrating metadata-specific features between branches. Experiments on publicly available ABIDE I and our self-collected dataset TJAD-ASD MRI reveal that the proposed MD-DGNN achieves the state of-the-art performance, demonstrating the superiority of our method.