Using the available cross‐sectional data of ENIGMA‐OCD, we studied medication effects in two additional ways. First, we used machine learning analysis of the previously analyzed cortical and subcortical measures from 2,304 OCD patients and 2,068 controls to assess whether anatomical group differences might be used to create a neuroimaging biomarker for OCD (Bruin et al., 2019). Classification performance across 10 different machine and deep learning approaches was limited, in this initial analysis that only used extracted measures derived from anatomical MRI. With site‐stratified cross‐validation, the receiver operating characteristic area under the curve (ROC‐AUC) ranged between 0.57 and 0.62, and the performance dropped to chance level (classification performance between 0.51 and 0.54) when leave‐one‐site‐out cross‐validation was used, indicating that these anatomical brain features, on their own, are not suitable as a biomarker for OCD. However, when patients were stratified according to whether they currently use medication, classification performance improved remarkably: Medicated OCD patients and healthy controls could then be distinguished with a 0.73 AUC (SD = 0.03, p corr < .001) (in contrast to unmedicated OCD and healthy controls, with 0.61
