How to remove or control confounds in predictive models, with applications to brain biomarkers.

Cross-validation that isolates nuisance effects gives an additional piece of information: confound-free prediction accuracy.