Large Language Models Are Poor Clinical Decision-Makers: A Comprehensive Benchmark

Abstract

The adoption of large language models (LLMs) to assist clinicians has attracted remarkable attention. Existing works mainly adopt the close-ended question-answering task with answer options for evaluation. However, in real clinical settings, many clinical decisions, such as treatment recommendations, involve answering open-ended questions without pre-set options. Meanwhile, existing studies mainly use accuracy to assess model performance. In this paper, we comprehensively benchmark diverse LLMs in healthcare, to clearly understand their strengths and weaknesses. Our benchmark containsseventasks andthirteendatasets across medical language generation, understanding, and reasoning. We conduct a detailed evaluation of existingsixteenLLMs in healthcare under both zero-shot and few-shot (i.e., 1,3,5-shot) learning settings. We report the results onfivemetrics (i.e. matching, faithfulness, comprehensiveness, generalizability, and robustness) that are critical in achieving trust from clinical users. We further invite medical experts to conduct human evaluation.