大型語言模型進行醫療診斷推理

Abstract

醫療診斷推理是臨床工作中的重要能力，它讓醫師能從病人身上「蒐集關鍵資訊」，並且利用蒐集到的資訊「預測診斷」。本論文以問診作為研究案例，探討大型語言模型之診斷推理能力，並提出進一步提升此推理能力之方法論。此項研究之主要貢獻有三： 一、發展大型語言模型角色扮演評估框架，用以評估大型語言模型之問診能力； 二、提出少樣本、零樣本之提示工程方法論，此方法論結合醫師診斷推理之思考過程，並以實驗佐證其能提升大型語言模型「蒐集關鍵資訊」以及「預測診斷」之表現。 三、顯示大型語言模型能透過儲存及擷取自生成之診斷推理過程，持續增進其診斷預測之能力。

Medical diagnostic reasoning is a crucial capability in clinical practice, enabling physicians to "collect key information" from patients and "predict diagnoses" based on the collected information. This thesis investigates the diagnostic reasoning abilities of large language models (LLMs) using history taking as a case study and proposes methodologies to further enhance these reasoning abilities. The main contributions of this research are threefold: (1) Development of an LLM Role-Playing Evaluation Framework to assess the history-taking abilities of LLMs. (2) Introduction of few-shot and zero-shot prompting methodologies that integrate the diagnostic reasoning processes of physicians, with experimental evidence demonstrating their effectiveness in improving LLMs' performance in ``collecting key information'' and ``predicting diagnoses''. (3) Showing that LLMs can continuously improve their diagnostic prediction capabilities through storing and retrieving self-generated diagnostic reasoning processes.