“…Various machine‐learning techniques have been applied to mass rs‐fcMRI data to objectively identify disorder‐specific abnormalities in mental disorders, with which automatic case–control classifications were employed. Table summarizes the previous attempts for schizophrenia, MDD, ADHD, and ASD . The take‐home messages of this summary are as follows: (i) irrespective of disorder type, classification accuracy is, overall, 80–90%, comparable to those based on structural MRI data; (ii) in many studies, especially for schizophrenia and MDD, the sample per group (case or control) is typically comprised of fewer than 100 participants; (iii) for all schizophrenia and MDD studies, the imaging data were acquired at a single site, whereas for many ADHD and ASD studies, the imaging data came from multiple sites, thanks to the recent multicenter imaging campaigns for these disorders; (iv) inter‐regional functional connectivity and the associated graph metrics are popular features used for classification; (v) head motions during scanning have been known to introduce artifacts in the functional connectivity estimate, the effects of which are controlled by regression, masking (scrubbing), or independent component analysis; (vi) BOLD signal fluctuations of non‐neuronal origins, such as respiration and cardiac activity, are removed by regressing out the signals in white matter and cerebrospinal fluid, although further inclusion of global signal fluctuation into the regressor is not unanimous among the studies due to the recent controversy; (vii) support vector machine (SVM) and its variants are popular prediction methods, although some studies use classifiers with embedded regularization frameworks, such as least absolute shrinkage and selection operator (LASSO); (viii) leave‐one‐out and k‐fold cross‐validation procedures are popular methods for model evaluation; and (ix) for all but one study, the generalization capability of a classification scheme is untested in an independent cohort.…”