Virginia Via Research Day Book 2026

Medical Student Research Clinical

DEVELOPING A MACHINE-LEARNING MODEL TO DETECT EMERGENCY DEPARTMENT MISTRIAGE USING STRUCTURED DATA, CLINICAL NOTES, AND HOSPITAL OUTCOMES 12

Nonnie Komon, BS; Fred Rawlins II, DO; Nick Buhl, BS; Taylor Daniels, MS Corresponding author: Nkomon@vcom.edu

VCOM-Virginia, Blacksburg, Virginia

Context: Emergency Department (ED) triage relies on rapid human judgment, but prior research shows that clinical decision-making can be influenced by time pressure, cognitive overload, and implicit bias. Studies have demonstrated that mistriage disproportionately affects certain demographic groups, contributing to delayed care and worsening outcomes. Under-triage can lead to inequitable care delivery, delay in care, and critical outcomes for patients. Over-triage can lead to overcrowding of the ED and misuse of resources. Prior studies have shown that traditional triage tools, such as the Emergency Severity Index (ESI) can be inaccurate, especially for high-risk symptoms and in vulnerable populations. Machine learning models integrating structured EHR variables and natural language processing (NLP) of triage notes have demonstrated improved detection of high-risk patients. Objective: To develop an outcomes-based modeling approach for identifying emergency department mis-triage using information available at the time of initial presentation.

Methods: We analyzed over 425,000 de-identified emergency department encounters from the publicly available MIMIC-IV database. Mis-triage was defined by discordance between the initial ESI triage level and downstream clinical outcomes, including ICU admission and in-hospital mortality. Structured variables such as vital signs were analyzed alongside information derived from triage notes using a combined modeling approach. Model performance was evaluated using standard classification metrics, and feature contributions were examined using SHAP based methods. Preliminary Results: The outcomes-based model demonstrated reliable identification of under triage during initial testing, achieving 86% accuracy for under-triage classification. Feature analysis suggested that physiologic measures such as oxygen saturation and heart rate, as well as high risk symptom descriptors documented in triage notes, contributed meaningfully to model predictions.

Conclusion: These preliminary findings suggest that an outcomes-based approach using early ED data may help identify potential mis-triage more accurately than reliance on triage assignment alone. Incorporating downstream clinical outcomes appears to improve the clinical relevance of mis-triage labeling from a modeling standpoint. Future work will focus on refining model performance and exploring applications for quality improvement in emergency department triage, including a clinician-facing interface. IRB pending: Package 2394644-1

Table of Contents

183 2026 Research Recognition Day