Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a tool-augmented large language model (LLM) agent designed for biomedical entity linking (BEL) that leverages the Unified Medical Language System (UMLS) to improve disambiguation accuracy. By combining lexical, semantic, and relational utilities from UMLS with a hybrid symbolic-neural framework, the agent achieves a top-1 exact CUI match of 68.3% on a benchmark dataset, significantly outperforming existing methods like QuickUMLS and BioBERT. The findings highlight the effectiveness of ontology-grounded tool use and agentic reasoning in enhancing precision and interpretability in BEL tasks.
A novel LLM agent utilizing UMLS tools achieves over three times the disambiguation accuracy of leading biomedical entity linking methods.
: Biomedical entity linking (BEL) aims to normalize textual mentions of medical concepts to standardized identifiers within the Unified Medical Language System (UMLS), which integrates over 190 vocabularies and 3.45 million Concept Unique Identifiers ( CUIs ). The scale and lexical variability of UMLS make BEL a persistent challenge. Existing lexical and neural methods achieve limited disambiguation accuracy and underuse the ontology’s structure. This paper presents a tool-augmented large language model (LLM) agent that performs self-directed disambiguation by orchestrating lexical, semantic, and relational UMLS utilities. The agent integrates SapBERT-based candidate retrieval with semantic-type validation, ontology-driven neighbor expansion, and iterative reasoning to refine concept selection. This hybrid symbolic–neural framework enables dynamic grounding and interpretable decision-making. On a 1,000-mention subset of the MedMentions ST21pv benchmark, the proposed system achieves 68.3% top-1 exact CUI match, outperforming QuickUMLS (20.5%), BioBERT (33.2%), GPT-4o single-stage few-shot prompting approach (35.1%), SciSpacy (39%), and SapBERT (51.8%). The study demonstrates that ontology-grounded tool use and agentic reasoning substantially improve BEL precision compared to static embedding or direct LLM inference. Contributions include (i) a unified agentic framework for UMLS-guided linking, (ii) integration of structured knowledge with LLM reasoning via function-calling tools, and (iii) an empirical benchmark establishing the benefits and trade-offs of tool-augmented BEL in accuracy, interpretability, and cost.