基於知識圖譜引導大型語言模型主動提問：應用於多輪交互式醫療推理

Abstract

大型語言模型（LLM）在醫療推理任務上已展現巨大潛力，但現有的研究大多是在資訊完整的理想化情境下進行，這與臨床實踐中資訊不完整的現實有所脫節。為解決此問題，本研究提出了一個新穎的互動式推理框架，旨在賦予LLM主動提問以收集關鍵資訊的能力。我們的核心方法是設計一個知識圖譜推理器（KG Reasoner），它能在專業的醫療知識圖譜中探索，以識別當前最重要的資訊缺口，從而為LLM的提問提供策略性指引。此外，我們提出一個受臨床鑑別診斷思維啟發的信心評估機制，讓系統能夠準確評估自身的不確定性，並在必要時觸發提問，而非做出草率的決策。 為了驗證我們方法的有效性，我們在一個互動式醫學問答基準上進行了實驗。實驗結果表明，相較於如MedIQ等現有的基準模型，我們的框架能透過更具策略性的提問，有效收集關鍵的病患資訊，並避免了因過早診斷關閉而導致的錯誤。在最終的問答準確率上，我們的模型取得了顯著的提升，證明了本研究所提出的KG引導的提問策略，是讓AI模型更貼近真實臨床實踐、提升其可靠性與安全性的可行途徑。

While Large Language Models (LLMs) have demonstrated significant potential in medical reasoning, existing research is often conducted under the idealized assumption of complete information, which contrasts with the reality of incomplete information in clinical practice. To address this gap, this thesis proposes a novel interactive reasoning framework designed to empower LLMs with the ability to proactively ask questions to gather critical information. Our core methodology involves a Knowledge Graph (KG) Reasoner that explores a professional medical KG to identify the most crucial information gaps, thereby providing strategic guidance for the LLM's questioning. Furthermore, we introduce a confidence estimation mechanism inspired by the clinical process of differential diagnosis, enabling the system to accurately assess its own uncertainty and trigger questions when necessary, rather than making premature decisions. To validate our approach, we conducted experiments on an interactive medical question-answering benchmark. The results demonstrate that, compared to existing baselines like MedIQ, our framework can effectively gather key patient information through more strategic questioning and avoid errors caused by premature diagnostic closure. Our model achieves a significant improvement in final question-answering accuracy, proving that the proposed KG-guided questioning strategy is a viable path toward making AI models more aligned with real-world clinical workflows, thereby enhancing their reliability and safety.