一句话：提出 MARACTUS，用 LLM 智能体把非结构化临床文档转换为机器可读行动树，支撑模型驱动工程与形式验证。
临床指南是标准化、循证诊疗的基础，但其叙事化、非结构化的形式容易导致不一致与解读错误；WHO 也强调需要机器可读的规约来支持临床决策，而现有此类规约多由人工编写，易出错。本文提出 MARACTUS：一种智能体式方法，将非结构化医疗文档转换为表示临床流程的结构化机器可读行动树（MAT）。生成的 MAT 可输入模型驱动工程流水线、支持形式验证，并集成到临床决策支持系统中。MARACTUS 在编排框架内组合了输入构造器、大语言模型与语法检查器，在保证遵循固定文法的同时，借助经策展的知识库保留临床医生的监督。作者以两个示例临床案例、两个先进 LLM，在不同文档子集上比较零样本与少样本提示策略进行验证：语法正确率达 97.5%，相对人工基准的临床成功率为 62.5%，且领域专家在多个评估维度上给其中过半 MAT 打出了 ≥4/5 的临床有效性评分。
Clinical guidelines are essential for standardized and evidence-based patient care, yet their narrative, unstructured form often leads to inconsistencies and interpretation errors. As emphasized by the WHO, there is a need for machine-readable specifications to support clinical decision-making, but these are currently developed by hand, making the process error-prone. We present MARACTUS, an agentic methodology for converting unstructured healthcare documents into structured Machine-readable Action Trees (MATs) that represent clinical procedures. The produced MATs can feed model-driven engineering pipelines, enable formal verification, and integrate into clinical decision support systems, maximizing their impact on safer, more transparent, and continuously updatable healthcare software. MARACTUS combines an input constructor, a Large Language Model (LLM), and a syntax checker within an orchestration framework, ensuring adherence to a fixed grammar while maintaining clinician oversight via a curated knowledge base. We validate MARACTUS using two exemplar clinical cases and two state-of-the-art LLMs, comparing zero-shot and few-shot prompting strategies across subsets of documents. Our evaluation demonstrates a syntactic correctness rate of 97.5% and a clinical success rate of 62.5% relative to human ground truth, while domain experts rated over half of the MATs as clinically valid (with scores ≥ 4/5) across multiple evaluation dimensions.
@inproceedings{paperbot12061,
title = {Agentic Generation of Structured Clinical Specifications for Digital Healthcare Services},
author = {Bruno Guindani and Matteo Camilli and Livia Lestingi and Marcello Maria Bersani},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786581.3786943}
}