Classification of Autism Spectrum Disorder Subtypes based on Graph Attention Network

The diagnosis and detection of autism spectrum disorder (ASD) have remained significant challenges for clinicians and researchers over the past several decades. Based on symptom severity and developmental profiles, specific clinical presentations, such as autism, Asperger’s, and pervasive developmental disorder—not otherwise specified (PDD-NOS), were previously defined. While numerous studies have explored the classification of ASD patients versus healthy controls, relatively few have addressed the more challenging task of detecting ASD subtypes, despite its clinical relevance for personalized treatment. With the goal of solving this challenge, we developed a graph attention network (GAT)-based model for the classification of ASD subtypes. The model incorporates two kinds of frequency-domain features, which have been validated in previous studies, along with the brain pattern features derived from tensor decomposition. Feature graphs were constructed based on inter-ROI distances to simulate the functional brain network structures. A 3-layer GAT was trained to perform ASD subtype classification, achieving 65.85% accuracy and a weighted-F1 score of 65.55%. Our results demonstrated the effectiveness of graph learning for ASD subtype classification and highlighted the potential of ALFF, fALFF, and brain pattern features as neuroimaging biomarkers for ASD subtype detection.