ML-driven clinical documentation

Abstract

Electronic health records (EHRs) have irrevocably changed the practice of medicine by systematizing the collection of patient-level data. However, clinicians currently spend more time documenting information in EHRs than interacting directly with patients, and have adapted to time-intensive note-writing by authoring free-text notes overloaded with jargon and acronyms. Clinical notes are therefore difficult to parse and largely unstructured. This negatively impacts the ability of EHR systems to convey information between different clinicians and institutions, to communicate medical findings to patients, and to allow for programmatic ingestion of data to derive further automatically-learned insights. In this thesis, we present a new EHR system that addresses these problems by using novel machine learning methods to streamline the processes by which clinicians enter in new information and surface relevant details from past medical records. Our intelligent interface aids physicians as they type, allowing for automatic suggestion and live-tagging of clinical concepts to alleviate documentation burden, while simultaneously enabling clinical decision support and contextual information synthesis. Furthermore, as clinicians craft notes we automatically structure and curate their free-text inputs, allowing for further data-driven innovation and improvement. This EHR can reduce physician burnout, decrease diagnostic error, and improve patient outcomes, all while collecting a corpus of clean, labelled clinical data. Our system is currently deployed live at the Beth Israel Deaconess Medical Center Emergency Department and is in use by doctors.