Knowledge-embedded large language models for emergency triage

Emergency departments (EDs) are crucial to healthcare but face persistent overcrowding. The Emergency Severity Index (ESI) triage system is vital for prioritizing patients based on acuity and resource needs but relies heavily on the subjective judgment of medical staff, leading to inconsistencies. This study developed a Sequential Domain and Task Adaptation (SDTA) framework for enhancing ED triage accuracy and consistency using large language models (LLMs). By training LLMs on clinical data and ESI-specific tasks, we significantly improved their performance compared to traditional prompt-engineered models, achieving accuracy levels comparable to or exceeding those of experienced emergency physicians. Notably, the fine-tuned models achieved high accuracy and perfect recall for high-risk cases. These findings highlight the potential of adapted LLMs to standardize triage decisions and reduce variability, thus offering a solution to alleviate overcrowding and enhance patient care outcomes.