Background: Accurate triage in prehospital emergency care is critical for improving patient outcomes but is often hindered by limited information, environmental constraints, and variability among providers. Although global interest in AI-assisted tria...
Background: Accurate triage in prehospital emergency care is critical for improving patient outcomes but is often hindered by limited information, environmental constraints, and variability among providers. Although global interest in AI-assisted triage is growing, evidence regarding the application of Large Language Models (LLMs) in prehospital settings particularly with the Pre-Korean Triage and Acuity Scale (Pre-KTAS) remains limited.
Objective: This study assessed the performance of ChatGPT, both before and after training with official Pre-KTAS materials, by analyzing its triage accuracy, consistency, agreement with emergency medical technicians (EMTs), patterns of over- and under-triage, and diagnostic performance (sensitivity, specificity, PPV, NPV, AUC). The findings provide foundational data for developing AI-assisted prehospital triage systems.
Methods: A total of 100 expert validated Pre-KTAS simulation scenarios were used to compare three evaluator groups: an untrained LLM, a Pre-KTAS trained LLM, and ten certified EMTs. Each scenario was evaluated 30 times by the LLMs in independent sessions. Statistical analyses included Welch ANOVA, ICC(3,1), Cohen’s kappa, weighted kappa, χ² tests, and ROC curve analysis.
Results: The trained LLM demonstrated significantly higher overall accuracy (0.81 ± 0.03) than the untrained model (0.69 ± 0.10, p<.001) and showed performance comparable to EMTs (0.80 ± 0.03). It achieved superior accuracy for Pre-KTAS levels 1, 4, and 5, and exhibited the highest consistency (ICC=0.940). Agreement with EMTs improved notably after training (κ=0.66 vs. 0.56, p<.001). The trained LLM showed reduced over and under triage and strong diagnostic performance (sensitivity 0.95, specificity 0.92, PPV 0.88, NPV 0.97). All groups demonstrated excellent discriminative ability (trained LLMAUC 0.934; EMTs 0.934).
Conclusion: Following targeted Pre-KTAS training, ChatGPT achieved triage performance comparable to or exceeding that of experienced EMTs, with high accuracy, reliability, and reduced misclassification rates. These results support the feasibility of integrating LLMs as decision-support tools in prehospital triage and provide essential groundwork for future AI-based prehospital emergency care systems.
Keywords: Prehospital triage, Artificial intelligence, Large Language Model, ChatGPT, Pre-KTAS, Emergency medical services