ChatGPT‐4 Consistency in Interpreting Laryngeal Clinical Images of Common Lesions and Disorders

• DOI: 10.1002/ohn.897
• Publication Date: 2024-07-24T10:01:34Z
• Authors: Antonino Maniaci; Carlos M. Chiesa‐...; Jérôme R. Lechien
• Keywords: Concordance

Abstract Objective To investigate the consistency of Chatbot Generative Pretrained Transformer (ChatGPT)‐4 in analysis clinical pictures common laryngological conditions. Study Design Prospective uncontrolled study. Setting Multicenter Methods Patient history and videolaryngostroboscopic images were presented to ChatGPT‐4 for differential diagnoses, management, treatment(s). responses assessed by 3 blinded laryngologists with artificial intelligence performance instrument (AIPI). The complexity cases between practitioners interpreting evaluated a 5‐point Likert Scale. intraclass correlation coefficient (ICC) was used measure strength interrater agreement. Results Forty patients mean score 2.60 ± 1.15. included. image interpretation 2.46 1.42. perfectly analyzed 6 (15%; 5/5), while GPT‐4 judges high 5 (12.5%; 4/5). Judges reported an ICC 0.965 ( P = .001). erroneously documented vocal fold irregularity (mass or lesion), glottic insufficiency, cord paralysis 21 (52.5%), 2 (0.05%), (12.5%) cases, respectively. indicated 153 63 additional examinations, respectively primary diagnosis correct 20.0% 25.0% cases. significantly associated AIPI r s 0.830; Conclusion is more efficient diagnosis, rather than analysis, selecting most adequate examinations treatments.