The Beth Israel Deaconess Medical Center Emergency Department in Boston is testing out a new EHR platform that addresses clinical documentation burden by using machine learning to offer auto-fill suggestions of terminology. It is a tool for data entry as well as data discovery. The system dynamically suggests relevant clinical concepts as a doctor drafts a note by leveraging features from both unstructured and structured medical data.
During a Sept. 8 presentation at the Harvard Clinical Informatics Lecture Series, Divya Gopinath, a graduate student in the MIT Computer Science & Artificial Intelligence Laboratory, described her team’s evolving efforts to create an intelligent interface that aids physicians as they type, while simultaneously decreasing documentation burden and enabling clinical decision support.
She noted that one problem the solution seeks to address is how much time clinicians spend documenting information in EHRs. She said they have adapted to this “time sink" by authoring jargon- and acronym-filled free-text notes that can be difficult to parse for humans and algorithms alike. She noted that with the advent of efforts such as Open Notes, patients can’t understand the medical jargon in these notes either.
The solution Gopinath and colleagues Monica Agrawal, Luke Murray, Steven Horng, David Karger, and David Sontag are developing seeks to address these problems by using novel machine learning methods to streamline data entry and exploration. The intelligent interface aids physicians as they type, simultaneously decreasing documentation burden, make documentation more meaningful and enabling clinical decision support. “This is still an ongoing project,” she said. “We are still interested in learning how doctors would want to use this.”
She demonstrated that as a physician types a note in the BIDMC emergency department, the new computer system features drop-down menus of terms to auto-complete the phrase they are typing. For instance, if they type the letters “sh,” the system might suggest shortness of breath. If the physician chooses that term, it is automatically added to the note as a structured concept. That eliminates keystrokes, and can be used to build prospectively labeled clinical data. To help create the ranked list of suggestions, the machine learning model uses any prior notes and vital signs.
The screen is divided into two parts: the left-hand side, where the documentation occurs, and the right-hand side, where data discovery happens. When concepts are tagged in the documentation, summaries of conditions and things to explore become available on the right side. For instance, as the doctor types and adds a term from the drop-down menu, a card pops up on the right-side sidebar that summarizes prior notes about the patient’s condition such as hypertension. The system also auto-fills sections of the documentation, such as medications related to the diagnosis and any past surgeries.
The system can also provide visualizations of quantitative trends, such as lab results over time. “We support aggregation of lab values over a chosen time frame like the last six months or year,” Gopinath said. She referred to some of these auto-fill aspects as “Easter Eggs”— meaning surprise benefits for clinicians. It alleviates redundant data entry from previous sections by pre-filling the medication section. The system also puts the patient’s diagnosis in context. When you add a tag such as Afib, it populates an Afib card for that patient that is contextualized with related vital signs and snippets of notes. Tagged terms bring up clinical decision support and lab values from the past.
As the team continues to refine the tool, Gopinath mentioned several next steps. One is to prompt doctors to ask about symptoms being present or absent to help build probabilistic models. Another is to make the system more dynamic and customized for individual users based on the keyboard shortcuts they use. She said research has been done that shows the auto-complete does significantly decrease the time spent writing notes.
A next step is evaluating the usefulness of the sidebar. Among the design considerations, Gophinath said, “Our tool should be opt-in. We don’t want to break someone’s work flow.”
