Purpose: An increasing number of residents use Large Language Models (LLMs) as adjunctive tools in their medical practice. This preliminary randomized controlled trial evaluated the performance of ChatGPT-4o as a text-based clinical decision support tool for first-year residents (PGY1) and compared the structure and completeness of ChatGPT-4o outputs with those of a board-certified otolaryngologist. Methods: PGY1 resident consultation outputs were randomly assigned to either ChatGPT-4o or a board-certified otolaryngologist for assessing the resident consideration of medical history, clinical presentation, and examination findings in establishing primary and differential diagnoses, management plan, and treatments. Two blinded board-certified otolaryngologists independently assessed ChatGPT-4o and otolaryngologist recommendations using the Artificial Intelligence Performance Instrument (AIPI). The Intraclass Correlation Coefficient (ICC) was used to measure interrater agreement. Results: Patient demographics were comparable between ChatGPT-4o (n = 40) and otolaryngologist groups (n = 40). Blinded assessment revealed comparable performance between the otolaryngologist and ChatGPT-4o for evaluating PGY1 resident consideration of medical history and clinical presentation for primary diagnosis, and management plan. ChatGPT-4o achieved higher AIPI scores for documenting differential diagnosis, completeness of additional examinations, and treatment explanations than the otolaryngologist. The mean AIPI score for ChatGPT-4o was significantly higher than the otolaryngologist score (14.20 ± 2.24 versus 12.34 ± 1.91; p = 0.001). Interrater reliability was adequate (ICC = 0.699). Discussion: To our knowledge, this is the first study comparing LLM-generated text-based decision-support feedback and human review for PGY1 residents. Future studies are needed to determine the role of AI as an adjunctive tool under human oversight. Implications for practice: ChatGPT-4o demonstrated comparable performance to a board-certified otolaryngologist for providing structured, text-based decision-support feedback to PGY1 residents, achieving higher performance for documenting differential diagnoses and treatment explanations reflecting greater informational completeness rather than superior clinical judgment.
Evaluation of ChatGPT-4o as a text-based clinical decision support tool for junior otolaryngology residents: A randomized controlled trial
Maniaci A.;
2026-01-01
Abstract
Purpose: An increasing number of residents use Large Language Models (LLMs) as adjunctive tools in their medical practice. This preliminary randomized controlled trial evaluated the performance of ChatGPT-4o as a text-based clinical decision support tool for first-year residents (PGY1) and compared the structure and completeness of ChatGPT-4o outputs with those of a board-certified otolaryngologist. Methods: PGY1 resident consultation outputs were randomly assigned to either ChatGPT-4o or a board-certified otolaryngologist for assessing the resident consideration of medical history, clinical presentation, and examination findings in establishing primary and differential diagnoses, management plan, and treatments. Two blinded board-certified otolaryngologists independently assessed ChatGPT-4o and otolaryngologist recommendations using the Artificial Intelligence Performance Instrument (AIPI). The Intraclass Correlation Coefficient (ICC) was used to measure interrater agreement. Results: Patient demographics were comparable between ChatGPT-4o (n = 40) and otolaryngologist groups (n = 40). Blinded assessment revealed comparable performance between the otolaryngologist and ChatGPT-4o for evaluating PGY1 resident consideration of medical history and clinical presentation for primary diagnosis, and management plan. ChatGPT-4o achieved higher AIPI scores for documenting differential diagnosis, completeness of additional examinations, and treatment explanations than the otolaryngologist. The mean AIPI score for ChatGPT-4o was significantly higher than the otolaryngologist score (14.20 ± 2.24 versus 12.34 ± 1.91; p = 0.001). Interrater reliability was adequate (ICC = 0.699). Discussion: To our knowledge, this is the first study comparing LLM-generated text-based decision-support feedback and human review for PGY1 residents. Future studies are needed to determine the role of AI as an adjunctive tool under human oversight. Implications for practice: ChatGPT-4o demonstrated comparable performance to a board-certified otolaryngologist for providing structured, text-based decision-support feedback to PGY1 residents, achieving higher performance for documenting differential diagnoses and treatment explanations reflecting greater informational completeness rather than superior clinical judgment.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
