Hop-Gated Graph Attention Network for ASD Diagnosis via PC-Based Graph Regularization Sparse Representation

Autism spectrum Disorder (ASD) is a neurodevelopmental disorder that severely affects the daily life of patients. Deep learning is widely used in the diagnosis of ASD. However, it is difficult to construct brain functional network (BFN) due to the locality and similarity of resting fMRI data. To solve this problem, we propose a PC-based graph regularization sparse representation to construct BFN by considering the local popular structure and the spatial prior structure of the brain. Then, the attention mechanism and phenotypic information are used to form a population graph and input into the proposed hop-gated graph attention network for diagnosis. The network considers the nodes that have no edge connection but provide important information, realizes the selective updating and forgetting of node information, and solves the oversmoothing phenomenon caused by the deepening of network layers. Through a large number of experiments on ABIDE data sets, compared with other BFN construction methods and graph neural network frameworks, the experimental results verify the effectiveness of our method.