LLM-based approaches for automated vocabulary mapping between SIGTAP and OMOP CDM concepts

In the context of global healthcare systems, integrating diverse medical terminologies and classification systems has become a priority due to the adoption of Electronic Health Record (EHR) systems and the imperative for information exchange between healthcare systems. This study addresses the necessity for mapping between the SIGTAP vocabulary used in Brazilian healthcare systems and the broader medical terms of the OMOP CDM terminologies. Two distinct pipelines are evaluated for the vocabulary mapping process, focusing on two subsets of the SIGTAP vocabulary: medicines and medical procedures. The first pipeline utilizes textual embeddings for semantic similarity evaluation, followed by Large Language Models (LLMs) for correspondences selection through a retrieval-augmented generation (RAG) approach. In the second pipeline, LLM agents employ predefined protocols for vocabulary mapping and query refinement. Our results show comparable performance between pipelines in both the Procedures subset (F₁ of 0.684 versus 0.678), and the Medicines subset (F₁ of 0.846 versus 0.839), indicating the viability of the multi-stage filtering approach. The second pipeline demonstrates an advantage over the first in terms of recall, highlighting the efficacy of dynamic query refinement by the agent. These findings provide evidence that LLM-based methods significantly reduce manual effort required by experts, enabling domain specialists to focus on more challenging cases.