An Electronic Health Record-Based Screening Tool to Support Safe Discharges of COVID-19 Patients in the Emergency Department

Emergency departments (EDs) experience a large volume of patients that present with a wide range of COVID-19 symptoms. Patients with COVID-19 often present to EDs early in their disease course, when symptoms are highly variable and the risk of future decompensation is highest. ED clinicians need to make fast decisions around patient dispositions, often with limited data. The large variation in outcomes--from uneventful recovery to hospitalization, intensive care unit admission, and death--has meant that some patients discharged from the ED will later return requiring higher levels of care. Thus, there is an urgent need to quickly assess patient risks and allow for a more informed decision-making process in the ED.

This research developed and validated a COVID-19 ED return screening tool, CERST, to help predict a COVID-19 patient’s risk of returning to the ED, and for those identified at risk, predict the risk of hospitalization and in-hospital mortality. The CERST was developed using natural language processing (NLP), machine learning (ML) algorithms, and predictive modeling.

The specific aims of the research were as follows:

• Iteratively develop a concept map using mixed methods, which serves as the ontology categorizing predictive factors for COVID-19 ED returns to inform ML model development.
• Develop and evaluate ML algorithms predictive of ED return risk for COVID-19 patients.
• Prospectively validate a COVID-19 ED return screening tool, CERST, using real-time data.

Researchers used a mixed methods approach to develop a concept map, or visual aid, to identify key factors associated with COVID-19 patients' return to the ED and subsequent morbidity or mortality. These factors guided the development and evaluation of ML algorithms aimed at predicting ED return risk for COVID-19 patients. To comprehensively detect ED returns, electronic health record (EHR) data from two health systems, MedStar Health and the University of North Carolina Health, along with regional health information exchange (HIE) data, were used.

The ML model matched existing models in predicting ED returns for all patients. However, when trained on structured EHR data, it outperformed all other models, indicating that leveraging such data can significantly boost ML’s predictive power in this domain. The utility of NLP on extracted clinical text was highlighted, with a simplified and promising performance. The research identified the most significant factors associated with the risk for ED return and morbidity or mortality among returns, creating the foundation for CERST using EHR data to support safe COVID-19 discharges in the ED setting and demonstrating the potential for faster clinical decision-making compared to existing tools. Findings showed that ML and NLP can leverage the data-rich environment of the ED to provide timely guidance to clinicians on whether admission to the hospital may be needed, with the potential to prevent complications and improve quality care for COVID-19 patients in a highly resource-constrained clinical setting. The research also highlighted the benefits of using HIE data to predict post-discharge events across multiple health systems in a service area and showcased the ML model’s generalizability and versatility in clinical settings.