The identification of which brain regions contributed most to the classification resulted from a multivariate analysis, and the localization of these regions should therefore be interpreted with caution. MVPA techniques typically result in better discriminative ability between groups compared to standard univariate analyses by taking the distributed nature of effects into account, but they do not provide inherent localization information (i.e., attributing effect sizes to individual ROIs) as all features used for prediction are considered as a whole. We derived individual feature importance from the RFC classifier using permutation-based inference to find brain features that contributed both significantly and consistently (across CV folds) to classification performance. Feature importance was derived from classifications using neuroimaging data after regression of covariates to avoid any undesirable effects on the interpretation of weights caused by non-imaging features. No feature importance obtained for the main classifications (OCD patients vs. HC) in either pediatric, adult or combined samples was statistically significant. This is likely due to the low classification performances obtained, suggesting that the features used are either too noisy or non-informative for main diagnosis predictions
