Grounding Medical Terms in MeSH RDF: Resolution Coveage and Retrieval Efficiency

Abstract

Introduction: Linked datasets built with semantic-web technologies are large and growing in number, which makes querying them expensive and makes prior knowledge of their content valuable for targeted retrieval.

Objectives: This paper addresses two coupled problems in the medical domain using Linked Data principles: grounding medical term mentions in the Medical Subject Headings (MeSH) vocabulary — resolving a surface mention to its concept — and mapping the resulting terms, including compound candidate terms (multi-word expressions that may contain two or more candidate terms), by their most specific broader descriptor.

Methods: We exploit the Medical Subject Headings (MeSH) controlled vocabulary published by the U.S. National Library of Medicine (NLM) as RDF, and its public SPARQL endpoint, to assign each input term the most specific broader descriptor(s) that subsume it in the MeSH polyhierarchy. Because a term may appear in several MeSH hierarchies, the method naturally returns one or more class labels. We then propose an extension that classifies compound terms absent from MeSH by decomposing them into components and computing their least common subsumer over MeSH tree numbers. We describe a three-stage pipeline (input validation, SPARQL interrogation over HTTP, and descriptor extraction), the auto-generated SPARQL query at its core, and an evaluation against the live MeSH endpoint.

Results:  Measured against a tree-number-walking alternative, the single direct query returns the same classes about four times faster (warm median 543 ms vs. 2178 ms) using one HTTP round-trip instead of three to five. On a constructed benchmark spanning four MeSH domains (cardiology, oncology, pulmonology, neurology), the LCS extension assigns a correct class to all 3570 intra-domain concept pairs and correctly abstains on all 8207 cross-domain pairs, whereas on a free-text clinical corpus it recovers none of eleven terms — results that delimit the method’s scope precisely: it classifies compounds built from indexed concepts, not arbitrary free-text phrases. On a standard disease-mention corpus (NCBI Disease), enhancing the lookup with synonyms, case variants, and abbreviation expansion raises resolution coverage from 1.5% to 48.8%, while classification remains reliable once a term is resolved.

Conclusions: Overall, the direct SPARQL-based method classifies medical terms in MeSH quickly and reliably, The main challenge remains resolving terms from real clinical text, and synonym and abreviation matching appear to be the most promising way to improve this.