JMIR Publications

ABSTRACT

Background:

Emergency department (ED) triage determines treatment priority and resource allocation. Artificial intelligence (AI), including machine learning (ML) and large language models (LLMs), has been increasingly applied to triage, but evidence remains heterogeneous.

Objective:

This scoping review mapped review-level and primary-study evidence on AI/ML/LLM for ED triage accuracy.

Methods:

Guided by the Joanna Briggs Institute and reported according to PRISMA-ScR and PRISMA-S, we searched multiple electronic databases for records published in recent years. Independent reviewers screened, selected, and extracted data. Quality was assessed using validated critical appraisal tools for primary studies and reviews. Findings were synthesized thematically.

Results:

This scoping review provides an evidence map of ML and LLMs for ED triage accuracy. Traditional ML models performed well on structured retrospective data, often improving accuracy and reducing over- and under-triage. LLM performance varied widely across models, versions, prompts, languages, and clinical settings, and no consistent superiority over experienced clinicians was demonstrated. Sensitivity for the highest-acuity category was limited, and over-triage was common, raising safety concerns. Pediatric evidence suggested that experienced nurses often outperformed AI, with additional risks of over-triage and missed high-acuity cases. Electronic decision support systems improved triage consistency but sometimes increased triage time. Most primary studies were retrospective, single-center, or simulation-based, with high risk of bias, and prospective multicenter validation, fairness audits, and patient-centered outcomes remain scarce.

Conclusions:

AI/ML/LLM triage is promising but immature. Current evidence does not support autonomous AI triage. AI should remain a human-supervised adjunct, beginning with lower-risk clinician-facing applications. Future research should prioritize prospective multicenter validation, patient-centered outcomes, standardized reporting, external validation, fairness audits, workflow integration, human–AI interaction, and continuous safety monitoring.