Transforming and evaluating the UK Biobank to the OMOP Common Data Model for COVID-19 research and beyond

Václav Papez, Maxim Moinat, Erica A. Voss, Sofia Bazakou, Anne Van Winzum, Alessia Peviani, Stefan Payralbe, Elena Garcia Lara, Michael Kallfelz, Folkert W. Asselbergs, Daniel Prieto-Alhambra, Richard J. B. Dobson, Spiros C. DenaxasJ. Am. Medical Informatics Assoc 2022 · DOI

TL;DR

This paper details the process of converting the massive UK Biobank dataset into the OMOP Common Data Model (CDM). This transformation enables federated research on COVID-19 and other diseases by standardizing diverse data types, making it easier to combine with other datasets for large-scale analyses.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for rapid, large-scale research using real-world data to inform public health policies. However, combining data from different sources is a significant challenge due to variations in data formats, terminologies, and collection methods. International federated analyses, which are crucial for understanding disease variability and rare events (like vaccine side effects), are hindered by this data heterogeneity. Common Data Models (CDMs) like OMOP aim to solve this by standardizing data, but applying them to complex, large-scale biobank studies with rich, multimodal data presents unique challenges. This paper addresses the problem of harmonizing the UK Biobank, a vast repository of genetic and phenotypic data, into the OMOP CDM to facilitate such federated research.

Methods

The core of this work involved developing and applying an Extract, Transform, Load (ETL) process to convert UK Biobank data into OMOP CDM version 5.3. This process involved:

  • Data Profiling: Using the OHDSI White Rabbit tool to understand the structure and content of the UK Biobank data.
  • Syntactic and Semantic Mapping:
    • Baseline Data: Transforming self-reported illnesses, medications, procedures, and demographic information. This involved creating custom mapping tables and leveraging the UK Biobank's Athena vocabulary for non-standard terms. The USAGI tool was used to suggest mappings from non-standard to standard OMOP concepts.
  • Electronic Health Records (EHRs): Mapping data from primary care, hospitalizations, cancer registrations, and mortality records. This involved using existing OMOP vocabularies and custom mappings for various terminologies like SNOMED-CT, CTV3, ICD-10, and OPCS.
  • Handling Unmapped Data: A significant portion of unmapped data was transformed into the OMOP Observation domain with a Concept ID of 0, preserving the original field ID for future reference.
  • Validation: Employing OHDSI tools like Achilles, Data Quality Dashboard (DQD), and CDMInspection to assess the quality, conformance, and completeness of the transformed data. Synthetic data was also used for initial ETL testing.

Datasets

  • UK Biobank: A large prospective longitudinal study of 500,000 individuals, containing genetic and phenotypic data, including electronic health records (EHRs) linked from primary care, hospitalizations, cancer registrations, and mortality data across England, Scotland, and Wales.
  • EHR Data Sources: Primary care data from EMIS, Vision, and TPP systems; hospital admission data; cancer registration data; and mortality data.
  • Clinical Terminologies: SNOMED-CT, Clinical Terms Version 3 (CTV3), International Classification of Diseases (ICD-10 and ICD-9), OPCS-3 and OPCS-4, Dictionary of Medicines and Devices (dm+d), and proprietary coding systems.

Key Insights and Ideas

  • The Power of Standardization: The OMOP CDM is a powerful tool for harmonizing diverse and complex datasets, making them interoperable for large-scale research. This is crucial for tackling global health challenges like pandemics.
  • Questionnaire Data is Tricky: Transforming self-reported data from questionnaires into a structured CDM format is significantly more challenging than mapping structured EHR data. This requires extensive custom mapping and careful consideration of how to represent nuanced information like negations or missing answers.
  • Federated Analysis is the Future: By making the UK Biobank OMOP-compliant, researchers can now easily integrate it into federated analyses with other OMOP-transformed datasets, accelerating discovery and enabling studies on rare events.
  • The Value of Bespoke Vocabularies: For unique datasets like UK Biobank, creating custom OMOP vocabularies is essential to capture the richness of the original data and ensure accurate mapping.

Key Results

  • Over 502,505 participants and 1.37 billion rows of data were transformed into the OMOP CDM.
  • High mapping coverage was achieved for most data types, with >97% coverage for SNOMED-CT and ICD-10, and ~89% for surgical procedures (OPCS-4).
  • Older terminologies (ICD-9, OPCS-3) showed slightly lower coverage (91% and 77%, respectively).
  • Prescription data achieved a high mapping accuracy of 98.7%.
  • The transformed dataset successfully identified 3,086 participants with COVID-19.
  • Validation checks showed good concordance for demographic, risk factor, and comorbidity data between the source and transformed datasets, with only a small number of discrepancies.

Limitations and Future Work

  • Questionnaire Data Challenges: The paper acknowledges that representing questionnaire data in the OMOP CDM remains a challenge, requiring further research to develop robust methods for handling negations, missing answers, and local coding systems.
  • Genomic Data: Genomic data (e.g., SNPs) cannot be directly integrated into the OMOP CDM, necessitating separate analytical approaches or future extensions of the model.
  • Observation Period Discrepancies: Similar to previous studies, discrepancies can arise due to the OMOP CDM's definition of an observation period, which may exclude historical medical events. The authors note that this has been recently revised to allow for events outside the observation period in certain use cases.
  • Future Work: The authors suggest that their findings and methodology can inform similar data harmonization efforts for other large-scale biobank studies, such as the "All of Us" and Million Veteran Programs in the US, to facilitate participation in OHDSI studies. The transformed UK Biobank resource is highlighted as a valuable tool for ongoing and future research.

Summary updated

BibTeX
@article{papez2022transforming,
  author = {Václav Papez and Maxim Moinat and Erica A. Voss and Sofia Bazakou and Anne Van Winzum and Alessia Peviani and Stefan Payralbe and Elena Garcia Lara and Michael Kallfelz and Folkert W. Asselbergs and Daniel Prieto-Alhambra and Richard J. B. Dobson and Spiros C. Denaxas},
  title = {{Transforming and evaluating the UK Biobank to the OMOP Common Data Model for COVID-19 research and beyond}},
  journal = {J. Am. Medical Informatics Assoc.},
  year = {2022},
  doi = {10.1093/jamia/ocac203},
}
Figure