At Atropos Health, we innovated an approach combining NLP models with Large Language Models (LLMs) to map medication text strings to RxNorm in Real World Data. This method achieved 99.2% accuracy in ingredient correctness on a billion-scale dataset, significantly reducing the NLP model’s error rate from 9% to 0.7%. Notably, this technique displayed a high agreement level (Cohen’s κ of 0.899) between its results and human experts, demonstrating its reliability and potential for broad application.

This approach led to a 98% reduction in mapping costs compared to manual processes, emphasizing its scalability and cost-effectiveness. This efficient method did not require model fine-tuning, training, extensive data exposure to the LLM, or the use of a clinically-focused LLM at all. The presentation will explore the technical challenges, cost analysis, and future prospects of this innovative approach.