On the Limits of Selective AI Prediction: A Case Study in Clinical Decision Making