How computers match medical terms in text to the right concepts in biomedical databases

What is it about?

Medical and clinical knowledge is mostly written as free text in articles, patents, electronic health records, and even social media posts. This makes it difficult to search, reuse, and connect information, because the same disease, drug, or gene can be described with many different names, abbreviations, or acronyms. Our article reviews how computer systems tackle this problem using a task called named entity linking (NEL). NEL systems first find important terms in text, such as diseases, chemicals, or genes, and then link each term to a precise entry in a biomedical knowledge organisation system (KOS) like UMLS, MeSH, or SNOMED CT. We analysed 102 research papers published between 2013 and 2024 to understand how these systems are built, what knowledge sources they use, and how the field has changed over the last decade. We describe the main families of methods (rule‑based, dictionary‑based, graph‑based, machine learning and deep learning, and more recent large language model approaches) and how they are evaluated on shared datasets. We also map which biomedical and clinical vocabularies and ontologies are most used, and highlight gaps, such as limited coverage for some sub‑domains and languages beyond English. Finally, we summarise common challenges (for example ambiguous abbreviations, missing concepts in ontologies, and noisy training data) and point to promising research directions.

Featured Image

Photo by Tim Wildsmith on Unsplash

Photo by Tim Wildsmith on Unsplash

Why is it important?

Reliable linking between clinical or biomedical text and standardised vocabularies is essential for many downstream applications, from clinical decision support and safety monitoring to large‑scale research using electronic health records and scientific literature. If a system picks the wrong concept for a term, it can mislead later analyses, bias models, and ultimately affect patient‑facing tools. ​ By systematically organising the last decade of research on biomedical named entity linking and its underlying knowledge organisation systems, our survey gives practitioners a clear overview of available methods, resources, and evaluation datasets. This can help researchers choose appropriate techniques, identify suitable ontologies and benchmarks, and avoid repeating known pitfalls. For ontology and vocabulary developers, our analysis highlights where current KOS fall short for NEL (for example in coverage, structure, and multilingual support), suggesting how they can evolve to better support text‑mining and AI applications. Ultimately, improving NEL helps turn unstructured biomedical text into high‑quality, computable data that can be shared, reused, and trusted.

Perspectives