Evaluation method where a model trains on all subjects except one, then tests on that held-out subject, measuring generalization across individuals.