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

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Terminology

Background:

• Extracting structured clinical information from unstructured free-text notes is a significant challenge.
• Current medical coding automation is limited, necessitating improved methods for clinical data utilization.

Purpose of the Study:

• To develop and evaluate a novel two-stage pipeline for linking clinical note spans to Systematized Nomenclature of Medicine-Clinical Terminology (SNOMED CT).
• To enhance the accuracy and robustness of automated medical concept extraction from electronic health records.

Main Methods:

• A two-stage pipeline combining fine-tuned sequence labeling for entity span detection and retrieval-augmented concept selection using a large language model (LLM).
• Utilized an instruction-tuned LLM for final concept selection after retrieving candidates from an embeddings database.
• Tested the pipeline on Medical Information Mart for Intensive Care (MIMIC-IV) discharge notes from the SNOMED CT Entity Linking Challenge.

Main Results:

• The proposed method demonstrated competitive accuracy in linking clinical entities to SNOMED CT codes.
• The system showed relative robustness in handling annotation ambiguities within clinical notes.

Conclusions:

• The developed two-stage pipeline offers a promising approach for automated clinical concept linking.
• This method can improve the extraction of clinically useful information from free-text clinical notes, advancing medical coding automation.