Natural Language Processing To Identify and Rank Clinically Relevant Information for EHRs in the Emergency Department
Project Final Report (PDF, 1.07 MB) Disclaimer
Disclaimer
Disclaimer details
Project Details -
Completed
-
Grant NumberR21 HS024541
-
AHRQ Funded Amount$297,844
-
Principal Investigator(s)
-
Organization
-
LocationAuroraColorado
-
Project Dates09/01/2016 - 08/31/2018
-
Care Setting
-
Medical Condition
-
Population
-
Type of Care
-
Health Care Theme
Given the density of data stored in an electronic health record (EHR), locating relevant clinical information in a timely fashion can be difficult, and may lead to a patient safety issue. Although problematic across all care settings, this challenge is heightened in the emergency department (ED), where providers may not have a prior relationship with a patient and often must make quick decisions about life-threatening conditions. Tools to enhance searching for information in EHRs are typically based on key words. Such searches are both inefficient and simplistic because they do not consider context and only capture exactly matched terms. In addition, existing search tools do not rank or raise the relevance of information for a given problem or complaint.
Newer search tools using natural language processing (NLP) and machine learning can be developed to automatically process, filter, and rank free-text information needed for a given patient’s complaint and present this alongside structured or coded information so that providers have a “snapshot” of relevant information. This project developed an NLP search tool that automatically identifies and ranks relevant clinical information based on a patient’s presenting complaint within the ED setting.
The specific aims of the project were as follows:
- Develop complaint-specific information elements.
- Develop and test an NLP-based information retrieval tool.
Seven subject matter experts (SME) in Emergency Medicine, Cardiology, and Orthopedics assisted in developing and refining the information elements to be extracted from the EHR based on chief complaints, and then evaluated the user interface that was developed. The relevant information elements for “chest pain” or “back pain” were informed by a literature review and two rounds of a modified Delphi method until all relevant items were identified and ranked. For a chief complaint of chest pain, 12 risk factors, 40 diagnoses, 20 diagnostic tests, 14 procedures, and 10 therapeutic drug classes were identified as relevant. For back pain, 13 risk factors, 25 diagnoses, 16 diagnostic tests, 4 procedures, and 5 therapeutic drug classes were identified.
Two years of structured and free text data were retrieved from the records of 9,347 patients presenting at the EDs at two academic medical centers with a complaint of chest or back pain. The data were used to develop and test the automated search tool. SMEs ranked the relevancy of information elements extracted manually for 99 patients and compared it to the NLP tool’s automated rankings. The SME manual ranking out-performed the tool’s methods on medications (92.2 to 98.2 percent versus 84.8 percent to 90.8 percent) but underperformed on problems (57.6 to 84.9 percent versus 71.7 to 88.7 percent). The results demonstrated that information retrieval methods using NLP and unsupervised machine learning can provide a reasonably accurate, low-effort, and scalable method for situation-specific clinical relevancy ranking.
Disclaimer
Disclaimer details