Learning the Joint Representation of Heterogeneous Temporal Events for Clinical Endpoint Prediction.