Fig 2. Schematic representation of the procedure for training the MDD classifier and evaluation of predictive power.

The MDD classifier was constructed using a nested cross-validation procedure in the discovery dataset. We also used undersampling and subsampling techniques. Generalization performances were evaluated by applying the constructed classifiers to the independent validation dataset. The machine learning classifiers are represented by PC cartoons. CV, cross-validation; HC, healthy control; MDD, major depressive disorder.