RxMap: an LLM-assisted tool for medication normalization

Objective: To develop a freely available, researcher-oriented system for accurate normalization of free-text medication strings to standardized RxNorm ingredient-level concepts. Methods: RxMap implements a fully automated normalization pipeline that combines deterministic RxNorm candidate generation with large language model (LLM)–assisted lexical parsing and hierarchical ingredient-level reconciliation. Raw medication strings are normalized to RxNorm ingredient-level (IN/MIN) concepts. ATC codes are assigned post-normalization, and an optional review interface supports transparent human review. Results: Evaluation on 22 624 unique medication strings from the IPUMS MEPS dataset demonstrated substantial improvements over deterministic RxNorm matching alone. The best-performing RxMap configuration improved RxCUI-level precision, recall, and F1-score from 0.865, 0.853, and 0.859 to 0.969, 0.964, and 0.966, respectively, with similar gains at the ingredient level. Conclusion: RxMap provides accurate, scalable normalization of free-text medication data to RxNorm concepts. The system operates fully automatically, offers optional review for quality control, and enables reproducible batch processing for research workflows.