Evaluating a Large Language Model for Emergency Triage Using Synthetic Clinical Dialogues: Toward a Multimodal...

Background/Objectives: Prioritizing patient safety through accurate emergency triage is critical during mass-gathering events such as Hajj and Umrah, where high patient volumes can significantly increase the risk of clinical errors. Although the Emergency Severity Index (ESI) provides a standardized triage protocol, it remains dependent on human judgment, while traditional artificial intelligence (AI) models often struggle to capture the nuanced clinical context embedded in patient narratives. To address this gap, this study evaluates GPT-5.2 using synthetically generated dialogues as a step toward a conceptual clinical decision-support framework that bridges structured electronic health records (EHRs) with triage interactions through clinical dialogues. By combining narrative inputs with structured physiological indicators, it explores the potential to support more informed clinical decision-making under controlled simulation conditions. Methods: To evaluate this contribution, we systematically assessed the role of clinical dialogues in triage performance using a balanced and stratified dataset of 1000 simulated Hajj-context scenarios. Results: The results demonstrate that incorporating clinical dialogues can significantly improve triage performance, with a 79.7% relative reduction in under-triage across all ESI levels. A more noticeable improvement was observed in high-acuity cases, with a 91.9% relative reduction in under-triage. In addition, the model achieved a high-acuity recall of 94.5%, highlighting its ability to identify critical "red-flag" symptoms that are often not captured in structured data. These findings suggest that narrative context may contribute to more accurate triage assessment and improved identification of high-acuity cases. Conclusions: Consequently, when evaluated within a controlled decision-support setting, this study suggests that large language model (LLM)-based systems may improve triage reliability and may contribute to safer triage decision-making in high-demand clinical environments.