Low predictive power of clinical features for relapse prediction after antidepressant discontinuation in a naturalistic setting