Predicting Autism Spectrum Disorder from Brain Imaging Data by Graph Convolutional Network

Abstract

Autism spectrum disorder (ASD) is a brain-based disorder characterized by social deficits and repetitive behaviors. Fast diagnostic prediction of ASD is important due to its prevalence: CDC estimates that 1 in 68 children has been identified with autism spectrum disorder. With the advancement in neuroimaging technology and AI, researchers have begun to build machine learning models that take the brain image of a patient as input, and predict whether he/she has ASD. A typical workflow is to preprocess a brain image into a network of connected brain regions, where indicative features are extracted using simple linear or convolutional models to be used for prediction. Recently, graph convolutional networks (GCNs) have become popular which can directly operate on graph data, such as the brain network. However, the recent work applying GCN for ASD prediction used a static population network which is not easy to use when the subject population updates. In this paper, we propose an ASD predictive model that applies GCN directly on a population-averaged brain network along with self-attention graph pooling, which can be easily applied to new patient diagnosis once trained, and it beats all existing models by a large margin in terms of accuracy (78% compared with the prior best, 70%, on the ABIDE-I database).