Accuracy of Information Provided by ChatGPT Regarding Liver Cancer Surveillance and Diagnosis
Jennie J. Cao, Daniel H. Kwon, Tahereh Ghaziani, Paul Y. Kwo, Gary Tse, Andrew Kesselman, Aya Kamaya, Justin R. Tse
Abstract
Accuracy of Information Provided by ChatGPT Regarding Liver Cancer Surveillance and DiagnosisChatGPT (OpenAI) is an artificial intelligence-based large language model (LLM) chatbot that has gathered significant interest in the medical community.It summarizes vast amounts of information to rapidly generate responses through a user-friendly interface.Potential medical use cases include creation of layperson medical ChatGPT reports and provision of screening test recommendations [1,2].Such applications are important because 60% of Americans seek medical information online [3].We evaluated the accuracy of information provided by ChatGPT relating to liver cancer screening and surveillance (hereafter, surveillance) and diagnosis.This retrospective study did not require institutional review board approval because it did not constitute human subject research.Twenty questions were created to address fundamental concepts in liver cancer surveillance and diagnosis, with attention to American College of Radiology LI-RADS and American Association for the Study of Liver Disease guidelines for hepatocellular carcinoma (HCC) [4,5].The questions were submitted to ChatGPT 3.5 (March 23 version) on April 9, 2023.Each question was submitted three times (yielding 60 total answers), consistent with a prior study [2].No feedback to ChatGPT was provided.ChatGPT's answers were independently assessed by six fellowship-trained physicians from three academic liver transplant centers who actively diagnose and/or treat liver cancer (see Supplemental Methods, available in the online supplement).The reviewers included two abdominal radiologists, two interventional radiologists, one medical oncologist, and one hepatologist.Each reviewer categorized each answer as accurate (score of 1; all information is true and relevant), inadequate (score of 0; all information is true, but either the information does not fully answer the question [i.e., missing