Counterfactual Phenotyping with Censored Time-to-Events