Medical Concept Normalization of German Clinical Expressions to SNOMED CT

Clinical narratives in electronic health records frequently contain clinical expressions describing medical conditions. Their free-text format limits interoperability and automated processing. Medical concept normalization (MCN) addresses this challenge by mapping textual expressions to standardized terminology concepts, such as those from SNOMED CT. However, clinical language is characterized by abbreviations, spelling variants, and short forms, which complicate automatic normalization.

In this work, we investigate the MCN of short German clinical expressions to SNOMED CT by comparing a direct large language model (LLM)-based normalization approach using GPT-5.4 with a hybrid retrieval approach that combines a medBERT.de bi-encoder for embedding-based retrieval and retrieval-augmented generation (RAG) reranking using GPT-5 variants.

The LLM-only baseline achieves a Recall@1 of 0.235, Recall@3 of 0.297, and Recall@5 of 0.303. In contrast, the embedding-based bi-encoder retrieval approach achieves a Recall@1 of 0.681, a Recall@3 of 0.783, and a Recall@5 of 0.812. Incorporating RAG-based LLM reranking further improves Recall@1 to 0.771, while Recall@3 and Recall@5 reach 0.809 and 0.812, respectively.

The results demonstrate that combining domain-specific embedding-based retrieval with LLM-based reranking substantially improves MCN performance compared to a standalone LLM approach. This hybrid strategy supports robust semantic matching between German clinical expressions and standardized terminology concepts, facilitating structured representation and enabling more effective secondary use of clinical text.