Machine learning for anxiety diagnosis using structural MRI does not generalize to unseen data: results from a large developmental cohort

This study explored whether structural magnetic resonance imaging (MRI) data from the Brazilian High-Risk Cohort for Mental Conditions (BHRCS) could be used to predict anxiety disorders in youth through machine learning. The sample included 209 participants with any anxiety disorder and 232 healthy controls, aged 6 to 14 years at baseline, followed across three waves from 2010 to 2019. Youth in the anxiety group had a diagnosis at any timepoint and valid MRI scans, while controls had no psychiatric diagnoses. Structural brain features extracted via FreeSurfer were used to train a random forest classifier to differentiate the two groups. The model incorporated feature selection, ComBat harmonization to address site effects, and 5-fold cross-validation. Performance was evaluated on a test sample and on a hold-out validation sample comprising 20% of the total dataset (n = 88). The model achieved an accuracy of 64% and an area under the curve (AUC) of 0.70 in the test set. However, performance in the validation sample was near chance, with an AUC of 0.51. These findings suggest that while structural MRI may capture some neurobiological differences in youth with anxiety, its predictive value for clinical classification remains limited in this context.