Title: Global multi-site benchmark classification of major depressive disorder using machine learning on cortical and subcortical features of 5,365 participants from the ENIGMA MDD dataset

Abstract: Machine learning (ML) techniques have gained popularity in the neuroimaging field due to their potential for classifying neuropsychiatric disorders based on brain patterns. However, the diagnostic predictive power of the existing algorithms has been limited by small sample sizes, lack of representativeness, data leakage, and/or overfitting. Here, we overcome these limitations and provide a generalizable ML classification benchmark of major depressive disorder (MDD, N=2,288) and healthy controls (HC, N =3,077) via Support Vector Machines (SVMs), based on brain measures from standardized ENIGMA analysis pipelines in FreeSurfer. First, splitting the dataset by age and sex allowed classification while balancing the distribution of age and sex in all cross-validation folds. Second, splitting the dataset by site allowed classification based on leaving sites out. Then, an SVM was trained and its performance was estimated on each test fold. Additionally, we stratified the data according to age of onset, antidepressant use, number of episodes and sex to see if less heterogenous groups would yield higher classification performance. Finally, we repeated the classification with harmonized dataset via ComBat. We were able to classify MDD patients vs HC with around 62% balanced accuracy and AUC of 0.65 when data were split into training and test sets by Age/Sex and without any correction for site effects. When the input features were corrected for site effects using ComBat methods and/or when splitting the data by Site, both accuracy and AUC dropped to approximately 52%. Similar results were observed in stratified groups. Other commonly used ML algorithms and harmonization techniques were exploratively evaluated and yielded similar results. Overall, we were not able to discriminate between MDD and HC via SVM using both cortical and subcortical brain features when accounting for site effects.