Abstract

Aim/Background: Triage is a critical gateway in frontline healthcare, managing patient flow and prioritizing clinical urgency. We evaluated the diagnostic accuracy, operational efficiency, and safety of Artificial Intelligence (AI)-assisted triage systems compared to traditional human-led triage across primary care and emergency settings.
Methods: Following PRISMA 2020 guidelines and registered in PROSPERO, we systematically searched PubMed, Embase, and the Cochrane Library for studies published between January 2019 and January 2026. Two independent reviewers screened articles comparing AI triage (Large Language Models [LLMs], neural networks, symptom checkers) against human clinicians. Risk of bias was assessed using the QUADAS-2 tool. A bivariate random-effects meta-analysis was performed to pool diagnostic accuracy metrics, with heterogeneity evaluated via the I^2 statistic and publication bias via Deeks' funnel plot.
Results: From 4,520 initial records, 20 core studies comprising 45,214 patient encounters were included. The pooled sensitivity for AI triage across all settings was 84.5% (95% CI: 79.2–88.7%), with a specificity of 81.2% (95% CI: 75.4–86.0%). Significant heterogeneity was observed (I^2 = 86%). Subgroup analysis revealed LLMs achieved higher sensitivity (88.1%) than rule-based symptom checkers (76.4%). A "safety paradox" was identified in high-acuity cases: advanced AI models exhibited a pooled under-triage rate of 12.0%, which was associated with adverse outcomes in emergency cohorts, compared to a 4.5% under-triage rate by human clinicians (p=0.03).
Conclusion: AI-assisted triage demonstrates comparable accuracy to human clinicians for routine presentations but presents safety risks in high-acuity scenarios due to elevated under-triage rates. A "human-in-the-loop" implementation is essential to mitigate these risks and safely optimize operational throughput

Key words: Keywords: Artificial Intelligence, Triage, Primary Health Care, Family Medicine, Diagnostic Accuracy, Patient Safety.