Towards Scalable Structured Data from Clinical Text

Abstract

The adoption of electronic health records (EHRs) presents an incredible opportunity to improve medicine both at the point-of-care and through retrospective research. Unfortunately, many pertinent variables are trapped in unstructured clinical note text. Automated extraction is difficult since clinical notes are written in their own jargon-heavy dialect, patient histories can contain hundreds of notes, and there is often minimal labeled data. In this thesis, I tackle these barriers from three interconnected angles: (i) the design of human-AI teams to speed up annotation workflows, (ii) the development of label-efficient modeling methods, and (iii) a re-design of electronic health records that incentivizes cleaner data at time of creation.