Large language models for generating script concordance test in obstetrics and gynecology: ChatGPT and Claude
Zuhal Yapıcı Coşkun1,2, Yavuz Selim Kıyak3, Özlem Coşkun3
1Department of Obstetrics and Gynecology, Ministry of Health, Ankara Bilkent City Hospital, Ankara, Türkiye.
Objective:
To evaluate the performance of large language models (ChatGPT-4o and Claude 3.5 Sonnet) to generate script concordance test (SCT) items for assessing clinical reasoning in obstetrics and gynecology.
Methods:
This cross-sectional study involved the generation of SCT items for five common diagnostic topics in obstetrics and gynecology in primary care settings. A total of 16 panelists evaluated the AI-generated SCT items against 11 predefined criteria. Descriptive statistics were used to compare the models' performance across criteria.
Results:
ChatGPT-4o had an overall agreement rate of 90.57% for SCT items meeting the quality criteria, while Claude 3.5 Sonnet achieved 91.48%. The criterion with the lowest scores was "The scenario is of appropriate difficulty for medical students," with ChatGPT-4o rated at 71.25% and Claude 3.5 Sonnet at 76.25%.
Conclusion:
Large language models can generate SCT items that effectively assess clinical reasoning; however, further refinement is required to ensure the appropriate level of difficulty for medical students. These findings highlight the potential of AI to enhance the efficiency of SCT generation in obstetrics and gynecology within primary care settings.
