Can we rely on generative AI for emergency patient triage classification?

Abstract

Triage involves classifying patients based on their need for treatment and the available medical resources. Owing to the rampant issue of Emergency Department (ED) overcrowding, medical sciences pursue faster and more consistent approaches to patient triage. Advances in generative artificial intelligence, particularly large language models (LLMs), have prompted interest in their potential use for clinical decision support. However, the humanness of these systems can foster human-like trust, which may overlook reliability and functional performance in favor of perceived competence and integrity. This study evaluates whether such trust is justified by examining the use of LLMs for emergency triage prediction. We compare LLM and expert triage decisions across 111 mass-casualty scenarios using the START/JUMPSTART and National Triage Scale (NTS) systems. Agreement between model predictions and expert solutions was evaluated using the Quadratic Weighted Kappa (QWK) statistic with expert-derived weight penalties. The results show moderate agreement between LLM predictions and expert triage decisions (QWK = 0.44), indicating that the tested model may not be sufficiently reliable for autonomous triage classification. An explainability analysis of the model responses suggests that while the model performs adequately in controlled scenarios, it struggles in typical mass-casualty situations, often prioritizing textual cues over triage assignment logic. These findings support calibrated trust in LLM-assisted triage and the continued need for human oversight. The flaws we have observed in the model reasoning pathways for triage decision-making make an argument for neuro-symbolic machine learning.