Multiscale Spatial–Temporal Graph Attention Network for fMRI Brain Disease Classification

Brain disease classification based on functional magnetic resonance imaging (fMRI) has become a hotspot in artificial intelligence research. Considering the graph structure properties of the brain network, graph learning methods have attracted increasing attention in recent work. However, these studies mainly construct spatial graph on a single scale to represent brain network, ignoring the complex functional interactions of brain network at multiple scales, as well as the potential of graph learning in temporal feature extraction. To address these issues, this study proposes a novel multi-scale spatial-temporal graph attention network (MSTGAT) for fMRI brain disease classification. We design multi-scale spatial graph attention learning (MS-GAT) and temporal graph attention learning (T-GAT) modules, in which the former one constructs multi-scale topological brain networks to learn and combine high-level spatial functional interactions among brain regions, and the latter one engages time encoding to learn temporal hyper-correlations among different time points. The learned spatial and temporal features are adaptively integrated, and the fused spatiotemporal features are incorporated into multi-layer perceptron for brain disease classification. Systematical experiments on three fMRI datasets indicate robust classification performance of our MSTGAT for different classification tasks and brain parcellations, outperforming several state-of-the-art classification methods. Our model demonstrates better classification accuracy and computational efficiency trade-off. We also identify important brain regions and connections associated with brain disease classification. Together, this study provides a promising model to effectively learn and integrate complementary features of multi-scale topological brain networks and spatiotemporal dynamics from the fMRI data to further promote brain disease classification.