Triage Accuracy and Error Patterns of Large Language Models Compared With 566 Emergency Care Professionals: Aggregated Data and SPSS Syntax

This record contains aggregated data and SPSS syntax for the study "Triage Accuracy and Error Patterns of Large Language Models Compared With 566 Emergency Care Professionals: A Scenario-Based Cross-Sectional Study". In the study, 566 emergency care professionals in six groups (emergency medicine specialists, emergency medicine residents, general practitioners, emergency department nurses, medical interns and emergency medical technicians/paramedics) and three large language models (ChatGPT, DeepSeek and Gemini) assigned a color-coded triage category to 60 written emergency department scenarios adapted from the Emergency Triage Education Kit, second edition. Triage_LLM_scenario_level.sav: one row per scenario (60 rows) with the a priori difficulty block, clinical domain, reference triage category, percentage of participants choosing the reference category overall and in each professional group, distribution of participant responses, modal responses, the category assigned by each model and its correctness, the majority vote of the models and the corrected item-total correlation. Triage_LLM_group_summary.sav: summary statistics by professional group and experience stratum (age, sex counts, emergency department experience, overall and category-specific accuracy, over-triage, under-triage and black responses, percentage of participants scoring above each model, and consistency on the eight repeated scenarios). Triage_LLM_scenario_analyses.sps: SPSS syntax for the scenario-level analyses (model accuracy, Cochran's Q test, McNemar tests, Spearman correlations and the Kruskal-Wallis test across difficulty blocks). Variable and value labels are stored in the SPSS data files. Participant-level responses are not included because participants were told that individual data would not be shared with third parties. The Kahramanmaras Sutcu Imam University Medical Research Ethics Committee approved the study (session 2025/17, decision no. 01, 12 May 2025).