<![CDATA[Triage accuracy remains suboptimal globally, particularly in developing countries employing simplified systems such as the Cipto Triage Method (CTM), and LLM performance on three-level triage systems in non-English contexts has not been evaluated. This study evaluated the diagnostic accuracy of four LLMs (DeepSeek V4 Pro, DeepSeek V4 Flash, Tencent Hy3, MiMo V2.5) for CTM classification as decision support tools for emergency department (ED) triage in resource-limited settings. The models were tested on a stratified random sample of 1,200 clinical records from the MIETIC corpus (v1.0.0; 5,528 unique cases after deduplication), relabeled from the Emergency Severity Index to CTM by four board-certified emergency physicians (400 cases per category), with physician consensus as the reference standard. Primary outcomes were overall accuracy and Cohen kappa; secondary outcomes included per-class sensitivity, specificity, positive and negative predictive value with 95% Wilson-score confidence intervals, undertriage detection rate, and API cost. Accuracy ranged from 66.5% to 68.1% (kappa 0.50–0.52). DeepSeek V4 Flash achieved the highest accuracy (68.1%, 95% CI 65.4–70.7%) at $0.207 total cost. No pairwise McNemar comparison was statistically significant (p > 0.05). All models exceeded benchmarks previously reported for ChatGPT on the Canadian Triage and Acuity Scale, while kappa values were comparable to GPT-4 on the Korean Triage and Acuity Scale. Undertriage detection varied markedly: MiMo V2.5 identified 56.6% of undertriaged cases versus 11.9% for DeepSeek V4 Pro. Lower-cost models performed equivalently to more expensive alternatives, though undertriage detection remained suboptimal across all models. These findings support the potential role of LLMs as adjunctive decision aids in resource-limited EDs, pending prospective validation in authentic clinical settings.]]>
Copyrights © 2026
