R. Andrew Taylor, M.D., M.H.S., Chris Chmura, R.N., Jeremiah Hinson, M.D., Ph.D., Benjamin Steinhart, M.S., Rohit Sangal, M.D., M.B.A., Arjun K. Venkatesh, M.D., M.B.A., M.H.S., Haipeng Xu, M.S., Inessa Cohen, M.P.H., Isaac V. Faustino, M.S., and Scott Levin, Ph.D.
Background: Emergency department (ED) triage methods are limited in their ability to risk-stratify patients and identify critical illnesses. Our objective was to examine the impact of implementing an artificial intelligence (AI)–informed ED triage tool on triage performance and patient flow.
Methods: We performed a multisite quality improvement study of an AI-informed triage clinical decision support (CDS) intervention. The CDS tool provided individualized triage recommendations and rationale based on the predicted risk of acute outcomes. Comparative analyses were conducted for 180 days preintervention versus postintervention, and across nurses exhibiting high and low levels of agreement with the CDS. Patient-centered performance outcomes and ED flow measures were examined and adjusted for potential confounders.
Results: A total of 174,648 ED visits (83,404 visits preintervention and 91,244 postintervention) were included across three EDs. The triage acuity distribution changed postintervention; low-acuity (level 4 or 5) visits increased 48.2% (23.9 to 35.4%), mid-acuity (level 3) visits decreased 18.7% (48.8 to 39.7%), and high-acuity (level 1 or 2) visits decreased 8.8% (27.3 to 24.9%). Following the intervention there was an increase in high-acuity identification (level 1 or 2 assignment) of patients requiring critical care (from 78.8% [95% confidence interval, 76.4 to 80.8%] to 83.1% [80.9 to 85.0%]; P<0.001) and no observed change in identification of emergency surgery postintervention (41.2% [37.8 to 43.7%] to 39.2% [36.4 to 41.5%]; P=0.16). Hospitalization rates among low-acuity patients were maintained (below 4%) postintervention. Variability in nurse agreement with the AI-informed triage CDS and triage performance was observed; the nurse subgroup with a high rate of agreement performed better than the AI independently, whereas those with lower rates of agreement performed worse. The AI triage CDS intervention was associated with reductions in median times from arrival to the initial care area (33.0%; 12.0 to 8.0 minutes), ED disposition (4.2%; 190.0 to 182.0 minutes), and to ED departure (6.1%; 311.0 to 292.0 minutes).
Conclusions: Implementation of an AI-informed triage CDS system was associated with improved triage performance and ED patient flow. Variability in nurse agreement with the AI triage CDS highlighted opportunities to harmonize AI and nurse judgment to further reduce unwanted variation in triage performance. (Funded by Beckman Coulter.)