Contrastive machine learning helps better characterize the neuroanatomical variation of ASD

Reviewed by Emily Henderson, B.Sc.Jun 2 2022

Understanding brain heterogeneity in people with autism spectrum disorder (ASD) could be pivotal to improving quality of life in affected individuals, by leading to more specific diagnoses and targeted behavioral interventions.

Now, application of a data-driven approach to MRI brain scans from individuals with ASD has revealed that – at least in terms of ASD-specific anatomical variation – ASD does not cluster into distinct subtypes but instead exists as continuous variation in brain structure. "Individual variation within ASD was better captured by continuous dimensions than by multiple distinct clusters," say the study's authors, "indicating that -; at least at the level of neuroanatomy -; dimensional approaches can provide a better account of individual variation than discrete diagnostic categories."

ASD is a prevalent neurodevelopmental condition that affects how people interact with others, communicate, learn and behave, and currently affects roughly 1 in 100 individuals worldwide. While extensive research efforts have been dedicated to understanding ASD's biological basis, progress towards these goals remains challenged because ASD is a highly heterogeneous condition – different people affected by ASD can differ in the severity of their behavioral symptoms, in their genetics and in underlying neuroanatomy. Researchers have debated whether individual differences in ASD are better understood as distinct subtypes or as variation along continuous dimensions.

To better characterize the neuroanatomical variation of ASD, Aidas Aglinskas and colleagues applied contrastive machine learning to MRI brain scans from the Autism Brain Imaging Data Exchange I (ABIDE I) dataset, compared to controls. The researchers were able to differentiate individual neuroanatomical variation specific to ASD from variation that characterizes the population as a whole; these patterns of variation were differentially associated with clinical and nonclinical participant characteristics, like age. (Notably, the authors were able to generalize these findings to a new and independent dataset, a property that facilitates its application in diagnostic settings.) However, rather than clustering around distinctive subtypes, Aglinskas et al. found evidence for continuous variation in brain structure that affect distinct sets of brain regions.