LLM-Based Extraction of Clinical Practice Guidelines into Structured Arguments

Clinical Practice Guidelines CPGs are the foundation of Evidence-Based Medicine but their long, complex and unstructured format makes them difficult to integrate in Clinical Decision Support Systems CDSS. While Large Language Models LLMs are great at reading text, their tendency to hallucinate and act as black boxes makes them unsafe for autonomous medical decisions. To solve this, we propose an automated pipeline that safely turns CPGs in raw PDF format into a structured, computable database of medical evidence expressed as formal arguments. First, we use computer vision to accurately extract complex tables and preserve the document's layout. Then, we constrain the LLM using strict clinical frameworks PICO and Toulmin to guarantee that every extracted claim is traceable and accurate. Finally, we use clustering and pruning methods to remove duplicate information and organize the data. The result is a clean, trustworthy knowledge base that lays the essential groundwork for formal argumentation graphs and reliable CDSS.