Outcome prediction from clinical text can prevent doctors from overlooking\npossible risks and help hospitals to plan capacities. We simulate patients at\nadmission time, when decision support can be especially valuable, and\ncontribute a novel admission to discharge task with four common outcome\nprediction targets: Diagnoses at discharge, procedures performed, in-hospital\nmortality and length-of-stay prediction. The ideal system should infer outcomes\nbased on symptoms, pre-conditions and risk factors of a patient. We evaluate\nthe effectiveness of language models to handle this scenario and propose\nclinical outcome pre-training to integrate knowledge about patient outcomes\nfrom multiple public sources. We further present a simple method to incorporate\nICD code hierarchy into the models. We show that our approach improves\nperformance on the outcome tasks against several baselines. A detailed analysis\nreveals further strengths of the model, including transferability, but also\nweaknesses such as handling of vital values and inconsistencies in the\nunderlying data.\n