UMLS::Similarity: Measuring the Relatedness and Similarity of Biomedical Concepts

Bridget McInnes, Ted Pedersen, Serguei Pakhomov, Ying Liu, Genevieve Melton-MeauxNAACL 2013

TL;DR

This paper introduces UMLS::Similarity, an open-source software tool that quantifies semantic similarity and relatedness between biomedical terms using the Unified Medical Language System (UMLS). It addresses the challenge of understanding complex relationships within vast biomedical vocabularies, offering a flexible and accessible solution for researchers.

What Problem Does This Paper Address?

Understanding the relationships between biomedical terms is crucial for various applications, from building better search engines to analyzing clinical records. However, the sheer volume and complexity of biomedical terminologies, like the Unified Medical Language System (UMLS), make it difficult to automatically determine how concepts are related. While some terms are clearly similar (e.g., "lung disease" and "Goodpasture's Syndrome" both being types of "disease"), others are related through different means, such as treatment or part-whole relationships (e.g., "Goodpasture's Syndrome" and "Doxycycline" as treatment, or "elbow" and "arm" as part-of). Existing methods often struggle to capture these nuanced relationships beyond simple hierarchical "is-a" structures. This paper aims to provide a comprehensive and accessible tool to measure both semantic similarity and relatedness within the UMLS.

Methods

UMLS::Similarity is built upon the foundation of WordNet::Similarity, adapting its measures for the unique structure of the UMLS. The core of the tool lies in its implementation of various semantic similarity and relatedness algorithms.

  • Similarity Measures: These quantify how alike two terms are, typically based on their proximity within an "is-a" hierarchy. The paper mentions measures like "path" (based on distance to a common ancestor) and others derived from established work.
  • Relatedness Measures: These go beyond hierarchical similarity to capture broader connections, such as "part-of" or "treats" relationships. These measures leverage information beyond the strict "is-a" structure, which the UMLS contains in abundance.
  • UMLS Integration: The software is designed to work with the UMLS Metathesaurus, a massive data warehouse that unifies over 100 biomedical and clinical data resources. It utilizes the Concept Unique Identifiers (CUIs) to represent concepts and their relationships.
  • Accessibility: UMLS::Similarity is available as a command-line interface, an API, and a Web interface, making it accessible to users with varying technical expertise. The Web interface, in particular, allows users to leverage the tool without the complex process of installing the UMLS itself.

Datasets

The primary dataset used is the Unified Medical Language System (UMLS). The paper specifically mentions using the 2012AA version of the UMLS for their experiments. The UMLS is a meta-database that integrates numerous other terminologies and ontologies, including:

  • Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT)
  • Foundational Model of Anatomy (FMA)
  • MeSH (Medical Subject Headings)
  • Gene Ontology (GO) (mentioned in the context of related work)

The paper does not introduce new datasets but rather leverages the comprehensive nature of the UMLS.

Key Insights and Ideas

  • Beyond "Is-A": The most significant insight is the clear distinction and importance of measuring relatedness in addition to similarity. Biomedical concepts are not just hierarchically related; they are connected through treatments, causes, symptoms, and part-whole relationships, which are crucial for practical applications.
  • UMLS as a Rich Resource: The UMLS, by integrating diverse terminologies, provides a much richer landscape for semantic analysis than single ontologies alone. UMLS::Similarity effectively harnesses this richness.
  • Accessibility is Key: Providing the tool through multiple interfaces (CLI, API, Web) significantly lowers the barrier to entry for researchers who need to perform semantic analysis on biomedical text.
  • Defaults Matter, But Customization is Powerful: While default settings (e.g., using MeSH with PAR/CHD relations) are provided for convenience, the ability to customize sources and relation types is essential for tailoring the analysis to specific research questions.

Key Results

The paper demonstrates the functionality of UMLS::Similarity through examples. For instance, when comparing "finger" and "arm":

  • Similarity (Path measure using MeSH hierarchy): A score of 0.25 was obtained.
  • Relatedness (Vector measure): A score of 0.5513 was obtained.

Table 2 and Table 3 in the paper provide detailed similarity and relatedness scores for "finger" and "arm" across different UMLS sources (FMA, SNOMED CT, MeSH) and relation types (PAR/CHD, CUI, RB/RN). These tables showcase how different sources and relation definitions can yield varying scores, highlighting the flexibility of the tool. The paper also notes that UMLS::Similarity has been used in various applications since its release, including health information search systems and disease similarity identification.

Limitations and Future Work

  • Definition Availability: A limitation acknowledged is that not all concepts in the UMLS have definitions, which can impact relatedness measures that rely on textual definitions.
  • Computational Cost: While not explicitly detailed as a limitation, the processing of the vast UMLS database for complex similarity and relatedness calculations can be computationally intensive.
  • Future Work: The paper implies that further development and exploration of new similarity and relatedness measures within the UMLS framework are ongoing. The tool's adaptability suggests it can serve as a platform for developing and evaluating novel semantic analysis techniques for biomedical data.

Summary updated

BibTeX
@inproceedings{mcinnes2013umls,
  author = {Bridget McInnes and Ted Pedersen and Serguei Pakhomov and Ying Liu and Genevieve Melton-Meaux},
  title = {{UMLS::Similarity: Measuring the Relatedness and Similarity of Biomedical Concepts}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2013},
}
Figure