Medication direction errors in pharmacies pose significant patient safety risks. This study explores using large language models (LLMs) with domain knowledge to reduce these errors. The MEDIC system, which emulates pharmacist reasoning, fine-tunes an LLM using expert-annotated data to extract core prescription components and assembles them into clear directions. Compared to LLM benchmarks, MEDIC showed significantly fewer near-miss events (errors caught before reaching the patient) in both retrospective and prospective evaluations (a 33% reduction in near-miss events during a production deployment). This demonstrates that LLMs, when combined with domain expertise and safety guardrails, can improve pharmacy accuracy and efficiency.

Cristobal Pais, Jianfeng Liu, Robert Voigt, Vin Gupta, Elizabeth Wade, Mohsen Bayati