Autism spectrum disorder diagnosis using the relational graph attention network

• Published in: Biomedical signal processing and control Vol. 85; p. 105090
• Main Authors: Gu, Xiaoai, Xie, Lihao, Xia, Yujing, Cheng, Yu, Liu, Lin, Tang, Lin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2023
• Subjects: ASD prediction; More diverse structural information; Relation-aware attention; Relation-aware attention; ASD prediction; More diverse structural information
• DOI: 10.1016/j.bspc.2023.105090
• ISSN: 1746-8094

•Define that each subject sample is a node, and each node corresponds to its own feature vector.•Construct an initial population graph using several information that may strongly influence the correlation between subjects of autism.•Combine the correlation of nodes in feature information and edges to obtain the importance of neighbor nodes to central nodes.•The attention coefficient is adopted by the relational graph attention convolutional layer to aggregate the information of neighbor nodes and complete the autism prediction for the designated subjects. Autism spectrum disorder (ASD) is a common neurodegenerative disorder, and its effective identification will facilitate medical diagnosis and treatment. Geometric deep learning methods, such as Graph Convolutional Neural Networks (GCN), have recently been proven to deliver generalized solutions for disease prediction. To enrich the valid information in ASD prediction, we explore various methods for constructing the population graph: Phenotype-Edge (P-Edge), fMRI-Edge (F-Edge) and phenotype combined with fMRI-Edge (PF-Edge). In addition, Graph Attention Networks (GAT) is introduced to capture correlation between subjects on graph’s node-features, which is ignored by previous GCN-based methods. However, the originally proposed architecture of GAT does not consider the edge-features. To exploit the structural information encoded in the edge-features, relation-aware attention is further introduced by Relational Graph Attention Network (RGAT) based on GAT. Based on three graph structures and RGAT, three ASD prediction models are proposed: RGAT involving P-Edge (p-RGAT), RGAT involving F-Edge (f-RGAT), and RGAT involving PF-Edge (pf-RGAT). GAT achieves an accuracy of 71.6% on the graph with only “site” and “sex” edge-features, but fails on the graph with more diverse edge-features. RGAT not only obtains stable predictions on different population graphs, but also learns more diverse edge-features while improving the accuracy by 1.4% compared to previous GCN. The further introduction of relation-aware attention through RGAT based on GAT gives the ASD prediction model the ability to learn more diverse information, while improves the model's generalization ability. This will facilitate the expansion of more valid structural information for the field of ASD prediction.