OHDSI Standardized Vocabularies—a large-scale centralized reference ontology for international data harmonization

Christian Reich, Anna Ostropolets, Patrick Ryan, Peter Rijnbeek, Martijn J. Schuemie, Alexander Davydov, Dmitry Dymshyts, George HripcsakJournal of the American Medical Informatics Association 2024 · DOI

TL;DR

The OHDSI Standardized Vocabularies provide a massive, open-source ontology system to harmonize patient data from diverse sources worldwide. This crucial resource enables large-scale, consistent observational research by standardizing clinical concepts, overcoming the challenges of fragmented and proprietary data representations.

What Problem Does This Paper Address?

Conducting large-scale observational research, especially across international borders, is hampered by the heterogeneity of electronic health record (EHR) and administrative claims data. Different healthcare systems use varying coding schemes, terminologies, and data structures, making it incredibly difficult to aggregate and analyze data consistently. While common data models (CDMs) like OMOP CDM provide a standardized structure, they require a robust, comprehensive, and efficient ontology to map the diverse source concepts to a unified representation. Existing resources like UMLS are too broad and complex for this specific purpose. This paper addresses the need for such a centralized, standardized vocabulary system to enable global data harmonization and facilitate reproducible, large-scale research.

Methods

The OHDSI Standardized Vocabularies are built as a comprehensive ontology that serves as a mandatory reference for all data sites within the OHDSI network. The process involves:

  • Importing and Consolidating: Incorporating numerous public standard vocabularies from various sources.
  • De Novo Generation: Creating new concepts and relationships where existing ones are insufficient.
  • Concept Standardization: Assigning a unique "standard concept" for each distinct clinical meaning, regardless of its original source vocabulary. Non-standard concepts are then mapped to these standard ones.
  • Domain Assignment: Categorizing each concept into specific clinical domains (e.g., Condition, Drug, Procedure, Measurement) that align with the OMOP CDM fields. This ensures that data is stratified correctly for analysis.
  • Hierarchical Structuring: Organizing concepts into polyhierarchies (allowing multiple parents) to represent complex relationships.
  • Quality Assurance: Implementing a multi-stage QA process involving automated and manual checks for data integrity, constraint enforcement, and semantic alignment.

Datasets

The paper describes the OHDSI Standardized Vocabularies as a system built upon and integrating data from numerous existing vocabularies. While not a "dataset" in the traditional sense for training a model, it is a curated collection of:

  • 136 Vocabularies: These are the source terminologies and coding schemes integrated into the OHDSI system.
  • UMLS: The Unified Medical Language System is a significant source, supporting 15 vocabularies within the OHDSI system.
  • External Sources: 101 vocabularies are incorporated from external sources.

The paper does not introduce new benchmarks but rather describes the creation and characteristics of the OHDSI Standardized Vocabularies themselves, which serve as a foundational resource for research.

Key Insights and Ideas

  • The "Closed-World" Assumption: OHDSI operates under a closed-world assumption for observational data, meaning all entities and facts are considered known. This influences how concepts are defined, particularly regarding negations or past events.
  • Standard vs. Source Concepts: The distinction between "standard concepts" (the canonical representation) and "source concepts" (the original term from the data) is critical for harmonization. Only standard concepts are used to represent facts in OMOP CDM tables.
  • Domain-Driven Stratification: Assigning concepts to specific clinical domains (Condition, Drug, etc.) is paramount. This ensures that, for instance, a procedure is always recorded in the "Procedure" field, regardless of its original vocabulary. This approach makes analysis independent of the specific vocabulary used in the source data.
  • The Challenge of Internationalization: While many source vocabularies are US-centric, the OHDSI network's international scope necessitates a vocabulary system that can accommodate global data, leading to the development of extensions like the RxNorm Extension.
  • Ontology Alignment is an Ongoing Endeavor: Standardizing vocabularies is not a one-time task. Concepts and relationships are constantly evolving, requiring continuous maintenance, review, and updates.

Key Results

  • Scale: As of March 2023, the vocabularies comprise 8,761,976 valid concepts (10,574,359 total) from 136 vocabularies.
  • Coverage: Standard concepts constitute 40.5% of the total valid concepts, while non-standard concepts make up 50.1%. Classification concepts account for the remaining 9.4%.
  • Relationships: Over 28 million valid relationships exist between concepts, with hierarchical relationships being the most common.
  • Uptake: The OHDSI Standardized Vocabularies are widely used, with hundreds of groups and several large data networks employing them. Over 8600 users have performed 50,000 downloads.
  • Impact: The system has enabled efficient phenotyping, covariate construction, patient-level prediction, population-level estimation, and standard reporting, addressing a major impediment to large-scale observational research.

Limitations and Future Work

  • Ongoing Maintenance: The authors acknowledge that standardizing vocabularies is a continuous process. Concepts and terms are constantly added, corrected, and relationships are updated.
  • Coverage Gaps: While extensive, there are still areas with insufficient coverage, particularly for detailed cancer diagnoses (ICD-O-3 is mentioned as an exception to SNOMED-CT's coverage) and emerging data types like genomic data.
  • Semantic Nuances: The system aims for comprehensive coverage but acknowledges that capturing all semantic nuances and relationships perfectly is challenging.
  • Future Work: OHDSI Working Groups are actively addressing shortcomings by adding new vocabularies, concepts, and relationships. The focus is on expanding coverage in areas like oncology and genetic data. The authors encourage researchers to exploit the resource and contribute their use cases, fostering a dynamic, community-driven evolution of the ontology.

Summary updated

Abstract

OHDSI has made available a comprehensive, open vocabulary system that is unmatched in its ability to support global observational research. We encourage researchers to exploit it and contribute their use cases to this dynamic resource.
BibTeX
@article{reich2024ohdsi,
  author = {Christian Reich and Anna Ostropolets and Patrick Ryan and Peter Rijnbeek and Martijn J. Schuemie and Alexander Davydov and Dmitry Dymshyts and George Hripcsak},
  title = {{OHDSI Standardized Vocabularies—a large-scale centralized reference ontology for international data harmonization}},
  journal = {Journal of the American Medical Informatics Association},
  year = {2024},
  doi = {10.1093/jamia/ocad247},
}
Figure