Abstract
Despite substantial efforts, it remains difficult to identify reliable neuroanatomic biomarkers of autism spectrum disorder (ASD) based on magnetic resonance imaging (MRI) and diffusion tensor imaging (DTI). Studies which use standard statistical methods to approach this task have been hampered by numerous challenges, many of which are innate to the mathematical formulation and assumptions of general linear models (GLM). Although the potential of alternative approaches such as machine learning (ML) to identify robust neuroanatomic correlates of psychiatric disease has long been acknowledged, few studies have attempted to evaluate the abilities of ML to identify structural brain abnormalities associated with ASD. Here we use a sample of 110 ASD patients and 83 typically developing (TD) volunteers (95 females) to assess the suitability of support vector machines (SVMs, a robust type of ML) as an alternative to standard statistical inference for identifying structural brain features which can reliably distinguish ASD patients from TD subjects of either sex, thereby facilitating the study of the interaction between ASD diagnosis and sex. We find that SVMs can perform these tasks with high accuracy and that the neuroanatomic correlates of ASD identified using SVMs overlap substantially with those found using conventional statistical methods. Our results confirm and establish SVMs as powerful ML tools for the study of ASD-related structural brain abnormalities. Additionally, they provide novel insights into the volumetric, morphometric, and connectomic correlates of this epidemiologically significant disorder.