医疗大语言模型因果知识图谱接地框架:心血管领域试点研究
文章背景与核心概要
当前的医疗决策支持大语言模型(LLM)评估往往过度侧重于单一答案的准确性,而忽视了对干预措施、底层机制、潜在危害、证据支持以及不确定性的严谨推理。这种局限性导致模型在面对复杂临床环境时,容易产生幻觉或缺乏可解释的医学依据。
为了解决这一痛点,本文作者提出了一种基于图结构的可复现评估框架,专门用于规范医疗场景下大语言模型面向干预措施的行为,并通过心血管领域的试点研究验证了其实用性。该框架核心包含四个部分:领域因果知识图谱、面向场景的子图提取、四种受控接地条件以及自动化标识符锚定评分流水线。试点结果表明,集成接地条件(C4)在因果边准确率、不良反应识别及证据准确性上表现最优,且大幅降低了无依据声明的发生率。
框架与执行摘要 (Executive Summary)
Current evaluations of Large Language Models (LLMs) in healthcare decision support heavily prioritize single-answer accuracy rather than rigorous reasoning regarding interventions, underlying mechanisms, potential harms, evidence, and uncertainty.
当前评估医疗决策支持中的大语言模型(LLM)时,往往严重偏向于单一答案的准确性,而忽视了对干预措施、底层机制、潜在危害、证据以及不确定性的严谨推理。
To address this limitation, the authors propose a reproducible, graph-centered evaluation framework for intervention-oriented LLM behavior in healthcare, demonstrating its utility via a cardiovascular pilot study. The framework comprises a domain causal knowledge graph, scenario-conditioned subgraph extraction, four controlled grounding conditions, and an automated identifier-anchored scoring pipeline.
为了解决这一局限性,作者提出了一种可复现的、以图为中心的评估框架,用于规范医疗领域中面向干预措施的 LLM 行为,并通过一项心血管试点研究展示了其效用。该框架由领域因果知识图谱、场景条件子图提取、四种受控接地条件以及自动化的标识符锚定评分流水线组成。
核心框架组件 (Key Framework Components)
- Domain Causal Knowledge Graph: Contains assertions structured as first-class, provenance-preserving nodes accompanied by stable identifiers.
- 领域因果知识图谱: 包含结构化的断言,这些断言作为一等、保留溯源的节点,并配有稳定的标识符。
- Scenario-Conditioned Subgraph Extraction: Automatically retrieves the relevant reified-assertion subgraph given any clinical scenario.
- 场景条件子图提取: 针对任意临床场景,自动检索相关的具体化断言子图。
- Controlled Grounding Conditions (C1–C4): Varies the method by which retrieved subgraphs are incorporated into the model's context:
- C1: Ungrounded
- C2: Knowledge-graph grounded
- C3: Causal-graph grounded
- C4: Integrated grounding
- 受控接地条件(C1–C4): 改变将检索到的子图纳入模型上下文的方法:
- C1: 未接地(Ungrounded)
- C2: 知识图谱接地(Knowledge-graph grounded)
- C3: 因果图接地(Causal-graph grounded)
- C4: 综合接地(Integrated grounding)
- Automated Scoring Pipeline: Anchored on assertion identifiers to compute intervention accuracy and other evaluation metrics on a single pass.
- 自动化评分流水线: 锚定于断言标识符,以在单次遍历中计算干预准确率和其他评估指标。
试点结果与发现 (Pilot Results & Findings)
To test the framework, the authors built a category-balanced scenario generator spanning eight distinct reasoning failure modes, instantiating it on a cardiovascular knowledge graph.
为了测试该框架,作者构建了一个涵盖八种不同推理失败模式的类别平衡场景生成器,并将其实例化在心血管知识图谱上。
The evaluation metrics successfully discriminated the grounding conditions along interpretable, non-redundant axes:
* Condition C4 (Integrated): Achieved the strongest causal edge F1 (0.838), adverse-effect F1 (0.833), evidence accuracy (0.738), and the lowest unsupported claim rate (0.114).
* Condition C1 (Ungrounded): Achieved the highest raw intervention accuracy (0.948), yet exhibited no measurable causal or evidential grounding.
评估指标成功地沿着可解释的、非冗余的轴线区分了各个接地条件: * 条件 C4(综合): 实现了最强的因果边 F1 值(
0.838)、不良反应 F1 值(0.833)、证据准确率(0.738),以及最低的无依据声明率(0.114)。 * 条件 C1(未接地): 获得了最高的原始干预准确率(0.948),但没有表现出任何可测量的因果或证据接地。
附加资源与链接 (Additional Resources & Links)
- Full-Text Access: View PDF | HTML Version | TeX Source
- License: Creative Commons Attribution 4.0 International

- External Citations: NASA ADS | Google Scholar | Semantic Scholar
- 全文访问: 查看 PDF | HTML 版本 | TeX 源码
- 许可证: 知识共享署名 4.0 国际版
- 外部引用: NASA ADS | 谷歌学术 | Semantic Scholar