Can medical students use artificial intelligence to learn transfusion? Evaluating <scp>ChatGPT</scp> responses to the American Society of Hematology medical student transfusion learning objectives

Abstract

BACKGROUND AND OBJECTIVES: Chat generative pretrained transformer (ChatGPT) is a large language model that is already in wide use among medical students as a means of learning. Many papers have evaluated ChatGPT as a presenter of medical knowledge for the general public and as a test-taking engine. For students who rely on ChatGPT to learn transfusion medicine, it is important to understand the limitations of the application. MATERIALS AND METHODS: Transfusion content from the American Society of Hematology 'medical student learning objectives' was edited into questions for the ChatGPT interface. The answers generated by ChatGPT were then marked by three experienced transfusion medicine physicians. RESULTS: ChatGPT scored on average 2.27 ± 0.21 on a 4-point scale. Two-thirds of its answers scored A, B or C, representing excellent, good or satisfactory achievement, respectively. One-third of ChatGPT's answers were assigned a failing grade. Simple questions of basic transfusion science performed the best; more complex questions as well as questions where clinical practice has evolved substantially over the last several years performed the worst. Some answers were assessed to be unsafe in clinical practice. CONCLUSION: As a resource for medical students learning transfusion medicine, ChatGPT has significant limitations. A considerable proportion of its answers to transfusion questions are unreliable, inaccurate and even unsafe. These incorrect answers are presented with the same authoritative tone as its correct answers, and an inexperienced learner would be challenged to differentiate between true and untrue responses. At the present time, it is not recommended for medical students to use ChatGPT to learn transfusion medicine.

Ähnliche Arbeiten