ChatGPT-Based Virtual Standardized Patient that Amends Overly Detailed Responses in Objective Structured Clinical Examinations

Naoki Shindo, Masaki Uto

Abstract: Objective structured clinical examinations (OSCEs) are a standardized examination for assessing medical and dental students. OSCEs involve a medical interview task in which examinees are evaluated based on their interactions with standardized patients (SPs), who are trained to respond according to specific clinical scenarios. However, preparing well-trained SPs incurs substantial costs. To overcome this limitation, the use of virtual SPs employing artificial intelligence has attracted considerable attention. In this study, we propose using ChatGPT to create virtual SPs capable of reacting to arbitrary clinical scenarios. However, the direct application of ChatGPT has drawbacks in that it tends to generate overly detailed responses, sometimes mentioning clinical information irrelevant to the examinees’ questions. Such behavior is unsuitable for the purpose of OSCEs. To address this limitation, we propose a mechanism that identifies and amends overly detailed responses from ChatGPT and integrates this mechanism into the ChatGPT-based virtual SP.