GRAPH CONVOLUTION MODEL FOR DIAGNOSING AUTISM SPECTRUM DISORDER USING DIVERSE GENETIC DATA

Abstract

Autism Spectrum Disorder (ASD) affects multiple domains of development, including social interaction, repetitive behavior, restricted interests, sensory processing, motor function, cognition, and emotional regulation. Early intervention can improve outcomes by addressing these challenges and supporting overall development. In this work, we developed a graph convolution model to distinguish children with ASD from non-autistic children using behavioral video recordings during psychiatrist-led intervention sessions and scores from the ISAA questionnaire completed by parents or caretakers. Behavioral videos were processed using a multi-channel 3D CNN to extract node features, while ISAA scores were used to compute similarity-based edges in the graph. We further integrated this behavioral graph with a previously constructed neuroimaging graph to obtain a more comprehensive representation of ASD. The proposed model achieved an accuracy of 86.34% using behavioral data alone and 87.56% after integrating graphs.