Seven prospective studies find that artificial intelligence can speed up and sharpen emergency department triage, while highlighting real‑world limits.

Key points

• Across seven real‑world studies, AI systems predicted triage levels with ~80%–99% accuracy, sometimes cutting documentation time.  
• One fuzzy‑logic tool hit 99% sensitivity and 99% specificity; a large machine‑learning protocol reduced life‑threatening mis‑triage from 1.2% to 0.9%.  
• Performance was uneven for the sickest patients, and evidence remains observational—researchers urge rigorous trials that track outcomes like ED length of stay, ICU admission, and mortality.  

A new systematic review of prospective clinical studies suggests artificial intelligence (AI) can support emergency department (ED) teams in triaging patients—potentially easing crowding and improving safety—while underscoring that AI should augment, not replace, experienced clinicians. The authors screened 1,633 papers and analyzed sevenprospective trials conducted in busy, real‑world EDs (see the PRISMA flow on page 4 of the paper).  

What the review found

Most studies layered AI onto familiar five‑level triage scales (like ESI, MTS, KTAS, and TTAS) and tested different approaches, from classic machine learning and deep learning to fuzzy logic. Reported triage‑prediction accuracy ranged from 80.5% to 99.1% across models. In head‑to‑head metrics, a fuzzy‑logic “Fuzzy Clips” system delivered ~99% accuracy with 99% sensitivity and 99% specificity, while a feed‑forward neural network achieved an F1 score of 72% and ~85% overall accuracy. (See Table 1–4, pages 5–10.)  

Several trials tracked operational impact. A voice‑enabled record system trimmed triage documentation by a median 27 seconds (204 sec vs 231 sec), though it left some fields incomplete. Another large study showed a machine‑learning triage protocol lowered “life‑threatening” mis‑triage from 1.2% to 0.9% compared with usual care. (See Table 1 and Table 4.)  

Not every tool performed well. In one trial, a symptom‑assessment app agreed with nurse triage only 33.9% of the time and over‑triaged 57% of cases—though 94.7% of uses were still considered clinically safe.  

Why this matters

Triage aims to quickly identify who needs care right now and who can safely wait. The review’s take‑home message: AI can be a useful second set of eyes, especially for mid‑acuity patients who make up most ED traffic, helping reduce over‑ and under‑triage and smoothing patient flow. But for the sickest patients (levels 1–2)—where a skilled nurse’s “first‑impression” assessment is vital—AI models either struggled or were excluded for safety reasons. (Discussion, pages 10–12.)  

Caveats and what’s next

The evidence base is still early. All included studies were prospective observational cohorts with small to moderate sample sizes, and reporting on statistics and sample‑size justifications was uneven. The authors call for randomized trials that use objective outcomes (e.g., ED length of stay, ICU admission, in‑hospital mortality) and for routine monitoring of model accuracy over time to prevent performance drift. (Quality appraisal and Limitations, pages 7, 11–12.)  

Bottom line for readers

AI triage tools are promising assistants: they can speed documentation, flag high‑risk patients, and reduce some errors. For now, they work best alongside trained ED staff—not in place of them—while researchers determine how to deploy them safely and fairly in everyday care.  

Source: Yi N, Baik D, Baek G. “The effects of applying artificial intelligence to triage in the emergency department: A systematic review of prospective studies.” Journal of Nursing Scholarship (2025). See PRISMA diagram on page 4 and performance tables on pages 5–10 of the article.  

Editor’s note: This article is for information only and is not a substitute for professional medical advice.