Clinical notes often describe the most important aspects of a patient's physiology and are therefore critical to medical research. However, these notes are typically inaccessible to researchers without prior removal of sensitive protected health information (PHI), a natural language processing (NLP) task referred to as de-identification. In order to build tools that perform deid, one typically needs the very same data that is private, thus creating a chicken-and-the-egg problem. In this work, we generate "fake" clinical notes where the deidentified information is replaced with real-seeming values (e.g. "Tim Lywood" instead of "George Beveridge") that still respect reasonable distributional semantics. We evaluate models trained on this synthetic data and show that they perform just as well as models trained on the sensitive PHI-bearing notes.
