A number of limitations deserve emphasis. First, we used a sample pooled from existing data across the world, without harmonized protocols for scanning, inclusion criteria or demographic and clinical characteristics. These sources of heterogeneity may limit classification performance, but this also provides an opportunity for model development using independent data sets and the discovery of biomarkers that are reproducible across study sites. Second, standardized FreeSurfer protocols were used for MR data processing to ensure reproducibility across sites. It has been shown that FreeSurfer tends to overestimate subcortical volumes in children48, and that MR field strength can affect regional cortical estimations49. However, these nonsystematic effects are expected to affect patients and HC equally and are therefore not expected to influence our results. Third, limited information on medication use was available. We were therefore only able to distinguish patients on antidepressants with or without adjuvant antipsychotics vs. those who had not received any medication. Medication history, medication dosage, and duration of use were unknown. Nonetheless, these coarsely defined medication groups enabled better case-control discrimination and good classification of medicated vs. unmedicated cases.
