The Utility of Large Language Models to Assist With Emergency Triage Decisions Within Otolaryngology.

Abstract

To determine whether contemporary large language models can match clinician performance in evaluating the urgency of emergency otolaryngology referrals. Blinded cross-sectional diagnostic reasoning study. Simulated emergency referral environment modeled on tertiary care otolaryngology practice. Thirty emergency referral scenarios spanning the spectrum of otolaryngologic urgency were independently evaluated by 4 large language models (GPT-5, GPT-4, DeepSeek, and Grok) and 4 clinicians (otolaryngology attending and resident, emergency attending and resident). Outputs were anonymized and scored by 10 blinded otolaryngologists for appropriateness of urgency and quality of explanation using a three-point scale. Statistical analyses included nonparametric group comparisons, adjusted ordinary least squares modeling with case-level control, and correlation of each entity's case profile with that of the otolaryngology attending. Inter-rater reliability was excellent. The otolaryngology attending achieved the highest overall performance. GPT-5 demonstrated comparable mean performance, with no statistically significant difference in either domain. GPT-4 scored modestly lower but received higher mean ratings than both emergency clinicians. DeepSeek and the otolaryngology resident demonstrated intermediate performance, while Grok and the emergency clinicians performed lowest. Group-level analyses showed no significant difference between the large language model and otolaryngology cohorts; both were rated higher than emergency clinicians in this sample. GPT-5 demonstrated triage performance comparable to the otolaryngology attending in this controlled sample. Large language models may support emergency decision-making and education when specialist consultation is limited, but require supervision, transparency, and local calibration.