Machine learning-integrated cortical feature analysis identifies cortical complexity as a discriminative feature in autism spectrum disorder

Autism spectrum disorder (ASD) is associated with heterogeneous cortical alterations, but which cortical properties most consistently contain ASD-related information remains unclear. We used machine learning as an integrative framework to identify convergent surface-based morphometric characteristics of ASD.

Five cortical measures (thickness, gyrification, cortical depth, fractal dimension, and surface ratio) from 100 Schaefer parcels were examined in 2075 ABIDE I and II participants (986 ASD, 1089 controls). Three linear and three nonlinear classifiers were evaluated with repeated 10-fold cross-validation, with confound regression and standardization confined to training folds. Permutation importance analysis was applied to identify important cortical features shared across models. An exploratory mixed-effects analysis in ASD examined age-dependent associations between these features and Autism Diagnostic Observation Schedule (ADOS) scores.

A radial basis function support vector machine performed best (mean balanced accuracy 0.607), and the advantage of nonlinear over linear models was modest and algorithm-dependent. Nonlinear models relied mainly on fractal dimension and linear models on cortical thickness, yet three features were shared by all six models, and the right ventral prefrontal parcel ranked first or second in every model. The default mode network (DMN) was the most represented in both families. Age moderated the association between this parcel's fractal dimension and ADOS social scores (corrected p = 0.039), with a positive association at younger ages.

Machine learning-integrated cortical feature analysis identified the ventral prefrontal DMN parcel as a convergent locus of ASD-related structural variation, captured by thickness in linear and fractal dimension in nonlinear classifiers.