Assessing the performance of ChatGPT in answering questions regarding cirrhosis and hepatocellular carcinoma

Abstract

AbstractBackgroundPatients with cirrhosis and hepatocellular carcinoma (HCC) require extensive and personalized care to improve outcomes. ChatGPT (Generative Pre-trained Transformer), a natural language processing model, holds potential to provide professional yet patient-friendly support.AimExamining the accuracy and reproducibility of ChatGPT in answering questions regarding knowledge, management, and emotional support for cirrhosis and HCC.MethodChatGPT’s responses to 164 questions were independently graded by two transplant hepatologists and resolved by a third reviewer. The performance of ChatGPT was compared to physicians or trainees in two validated questionnaires. ChatGPT’s knowledge on cirrhosis care was tested using 26 quality measures of cirrhosis management. Finally, its emotional support capacity was tested.ResultsChatGPT regurgitated extensive knowledge of cirrhosis and HCC, but for questions with correct responses, only a small proportion was labelled as comprehensive. The performance was better in basic knowledge, lifestyle, and treatment than in the domains of diagnosis and preventive medicine. For the quality measures, the model answered 76.9% of questions correctly but failed to specify decision-making cut-offs and treatment durations. Compared to physicians/trainees, ChatGPT lacked knowledge of regional guidelines variations, such as HCC screening criteria. However, it provided practical and multifaceted advice to patients and caregivers regarding the next steps and adjusting to a new diagnosis.ConclusionIn summary, we analyzed the areas of robustness and limitations of ChatGPT’s responses on the management of cirrhosis and HCC and relevant emotional support. ChatGPT may have a role as an adjunct informational tool for patients and physicians to improve outcomes.Conflict of Interest DisclosuresNone declared.Funding/SupportNoneEthics ApprovalSince all responses from ChatGPT were publicly available, approval from the institutional review board was not sought.Guarantors of article:Dr. Kuo and SpiegelAuthor ContributionsConcept and design: Yeo, Samaan, Spiegel, KuoAcquisition of data: Yeo, Samaan, Ng, VipaniData review: Ting, Trivedi, KuoStatistical analysis: YeoDrafting of the manuscript: Yeo, Samaan, NgCritical revision of the manuscript: All authors