Diagnosis of autism spectrum disorder (ASD) currently relies on behavioral observations because brain markers are unknown. Machine learning approaches can identify patterns in imaging data that predict diagnostic status, but most studies using functional connectivity MRI (fcMRI) data achieved only modest accuracies of 60-80%. We used conditional random forest (CRF), an ensemble learning technique protected against bias from feature correlation (which exists in fcMRI matrices). We selected 252 low-motion resting-state functional MRI scans from the Autism Brain Imaging Data Exchange, including 126 typically developing (TD) and 126 ASD participants, matched for age, nonverbal IQ, and head motion. A matrix of functional connectivities between 220 functionally defined regions of interest was used for diagnostic classification. In several runs, we achieved accuracies of 92-99% for classifiers with >300 features (most informative connections). Features, including pericentral somatosensory and motor regions, were disproportionately informative. Findings differed partially from a previous study in the same sample that used feature selection with random forest (which is biased by feature correlations). External validation in a smaller in-house data set, however, achieved only 67-71% accuracy. The large number of features in optimal models can be attributed to etiological heterogeneity under the clinical ASD umbrella. Lower accuracy in external validation is expected due to differences in unknown composition of ASD variants across samples. High accuracy in the main data set is unlikely due to noise overfitting, but rather indicates optimized characterization of a given cohort.
Keywords: autism spectrum disorder; conditional random forest; diagnostic prediction; intrinsic functional connectivity; machine learning; resting-state fMRI.