Continuity of care is crucial to ensuring positive health outcomes for\npatients discharged from an inpatient hospital setting, and improved\ninformation sharing can help. To share information, caregivers write discharge\nnotes containing action items to share with patients and their future\ncaregivers, but these action items are easily lost due to the lengthiness of\nthe documents. In this work, we describe our creation of a dataset of clinical\naction items annotated over MIMIC-III, the largest publicly available dataset\nof real clinical notes. This dataset, which we call CLIP, is annotated by\nphysicians and covers 718 documents representing 100K sentences. We describe\nthe task of extracting the action items from these documents as multi-aspect\nextractive summarization, with each aspect representing a type of action to be\ntaken. We evaluate several machine learning models on this task, and show that\nthe best models exploit in-domain language model pre-training on 59K\nunannotated documents, and incorporate context from neighboring sentences. We\nalso propose an approach to pre-training data selection that allows us to\nexplore the trade-off between size and domain-specificity of pre-training\ndatasets for this task.\n