CUICurate: A GraphRAG-based Framework for Automated Clinical Concept Curation for NLP applications

A research team led by Victoria Blake has introduced CUICurate, a novel GraphRAG framework that automates the labor-intensive process of clinical concept curation for medical NLP applications. The system addresses a critical bottleneck in healthcare data processing: mapping free text to Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs) and organizing them into clinically meaningful concept sets comprising related synonyms, subtypes, and supertypes.

The technical approach combines knowledge graph construction with large language model reasoning. Researchers first built a UMLS knowledge graph and embedded it for semantic retrieval. For each target concept, candidate CUIs were retrieved from the graph, followed by LLM filtering and classification steps comparing two models: GPT-5 and GPT-5-mini. Evaluation across five lexically heterogeneous clinical concepts revealed that CUICurate produced concept sets substantially larger and more complete than manual benchmarks while maintaining human-level precision. Notably, GPT-5-mini achieved higher recall during filtering, while GPT-5 produced classifications that more closely aligned with clinician judgments.

This breakthrough matters because clinical concept set curation has traditionally been labor-intensive, inconsistently performed, and poorly supported by existing tools. The framework offers computational efficiency with stable outputs across repeated runs, making it suitable for integration into clinical NLP pipelines. By automating what was previously manual work, CUICurate enables more scalable and reproducible approaches to medical data organization, particularly valuable for phenotyping and analytic requirements in healthcare research and applications.