Learning Effective Embeddings from Medical Notes

With the large amount of available data and the variety of features they offer, electronic health records (EHR) have gotten a lot of interest over recent years, and start to be widely used by the machine learning and bioinformatics communities. While typical numerical fields such as demographics, vitals, lab measurements, diagnoses and procedures, are natural to use in machine learning models, there is no consensus yet on how to use the free-text clinical notes. We show how embeddings can be learned from patients’ history of notes, at the word, note and patient level, using simple neural and sequence models. We show on various relevant evaluation tasks that these embeddings are easily transferable to smaller problems, where they enable accurate predictions using only clinical notes.