How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection