Automated Information Extraction Pipeline for Constructing ED Knowledge Graphs from Korean Clinical Problem Lists

Emergency department problem lists remain unstructured despite documenting major medical problems. We developed an information extraction pipeline for constructing ED knowledge graphs, comparing three normalization strategies and five LLMs on 250 annotated Samsung Medical Center ED problem lists. Llama 3.1 70B with Named Entity Disambiguation achieved optimal performance (Exact F1: 0.65, Partial F1: 0.70), improving F1 by 6.56% over raw text. The resulting knowledge graph from 2,000 problem lists contains 1,983 entities and 3,208 relations, establishing a practical framework for large-scale ED knowledge graph construction.