Cosine Similarity Graph Attention Networks for Autism Spectrum Disorder Diagnosis

Autism Spectrum Disorder (ASD) has been a long-standing human disorder, especially prevalent in children who do not yet have good communication skills, and the prevention and diagnosis of ASD is challenging. This has resulted in a variety of neural networks to help doctors intelligently assist in prediction, and graph neural networks (GNNs) have been highly successful in this task. The construction of graph structures, an extremely important component of graph neural networks, varies in the way they are constructed and makes a difference in disease prediction performance. In this paper, the cosine similarity method is used to address the problem of non-uniform data metrics due to differences in acquisition conditions between subjects. In order to solve the problem of neuronal death, the most commonly used activation function is replaced by elu, which is used and a higher classification accuracy is obtained, and the residual strategy is used so that the model can better learn the linear transformation in the task and enhance the model expression ability. Methods for fusing different phenotypic information to encode edge features were compared, where the best performance was improved by 3.1% compared to classical graph neural networks. New perspectives are provided for multi-site data analysis, and disease diagnosis.