A Two-Stage Pipeline for Linking Clinical Notes to SNOMED CT

Extracting clinically useful information from free-text notes remains challenging due to their unstructured nature, while medical coding is still only partially automated. We present a two-stage pipeline for linking spans in clinical notes to Systematized Nomenclature of Medicine-Clinical Terminology (SNOMED CT) that combines fine-tuned sequence labeling with retrieval-augmented concept selection. Stage 1 detects entity spans; Stage 2 retrieves candidates from an embeddings database and selects the final concept with an instruction tuned large language model (LLM). The proposed method has been tested in the SNOMED CT Entity Linking Challenge, which provided Medical Information Mart for Intensive Care (MIMIC-IV) discharge notes annotated with SNOMED CT codes. Results indicate competitive accuracy and relative robustness to annotation ambiguity.