Attention-based Graph Convolution Network for Autism Spectrum Disorder Identification

Autism Spectrum Disorder (ASD) is a brain development disorder that affects the social skills of patients. The diagnosis of ASD relies on the subjective judgment of the doctor about the patient's behavior. Therefore, it is imperative to develop objective diagnostic methods. Recently, many studies only focus on brain functional connectivity, while ignoring differences in phenotypic information between subjects and the impact of different representations. We propose an attention-based graph convolutional network (GCN) for ASD identification. The GCN is considered to be successful in modeling graph structures, and the attention mechanism allows the model to focus on task-related representations. In this work, we use the fMRI images of the brain to form the nodes, and use the attention mechanism to extract the phenotypic information to form the edges. Then, we utilize the GCN to learn representational information and the attention mechanism to flexibly fuse the representations. Finally, the multilayer perceptron is adopted to diagnose ASD. Our proposed method achieves an average accuracy of 88.54% and an average AUC of 90.04% on the ABIDE dataset.