Abstract: To address the problems of inefficiency, inconsistent standards, and difficulties in deep logical review inherent in traditional manual medical record review, an intelligent medical record review assistant based on Large Language Models (LLMs) was proposed. The method first employed a divide-and-conquer (DaC) strategy to construct a three-stage 'structuring - parallel review - aggregation' visual workflow on the Dify platform, thereby decomposing complex review tasks. Second, during the parallel review stage, prompt engineering techniques such as Expert Mimicry and Chain-of-Thought were utilized to guide multiple LLM nodes in performing concurrent, independent, and in-depth analysis across four dimensions: medical terminology, content completeness, diagnostic rationale, and treatment plan appropriateness. Finally, the review results from each dimension were aggregated to generate a unified, structured review report. Testing on both positive and negative medical record samples demonstrated that the assistant could accurately identify and locate predefined errors, generating a review report that included clear problem descriptions, judgment bases, and modification suggestions. This method provides an effective technical path for achieving efficient, accurate, and interpretable automated medical record review, and holds significant practical value for ensuring medical safety and improving the quality of healthcare services.

Key words: Large Language Model (LLM), medical record review, prompt engineering, Divide-and-Conquer (DaC), multi-dimensional parallel review

摘要： 针对传统人工病历审核存在的效率低下、标准不一及深度逻辑审查困难等问题，提出一种基于大语言模型的智能病历审核助手。该方法首先采用分而治之策略，在Dify平台上构建了“结构化-并行审核-聚合汇总”三阶段可视化工作流，将复杂的审核任务分解；其次，在并行审核阶段，运用专家模仿、思维链等提示工程技术，引导多个大语言模型节点从医学术语、内容完整性、诊断合理性和治疗方案合理性四个维度进行并发、独立的深度分析；最后，将各维度审核结果进行聚合，生成统一的结构化审核报告。对病历正负样本进行测试后的结果显示，该助手能够精准识别并定位预设错误，并生成包含清晰问题描述、判断依据及修改建议的审核报告。该方法为实现高效、准确且可解释的自动化病历审核提供了有效的技术路径，对保障医疗安全和提升医疗服务质量具有重要的实践价值。

关键词: 大语言模型, 病历审核, 提示工程, 分而治之, 多维度并行审核