The Use of Primary Care Big Data in Understanding the Pharmacoepidemiology of COVID-19: A Consensus Statement From the COVID-19 Primary Care Database Consortium

Hajira Dambha‐Miller, Simon J. Griffin, Duncan Young, Peter Watkinson, Pui San Tan, Ashley Kieran Clift, Rupert Payne, Carol Coupland, Jemma C. Hopewell, Jonathan Mant, Richard M. Martin, Julia Hippisley–CoxThe Annals of Family Medicine 2021 · DOI

TL;DR

This paper addresses the urgent need for standardized methods to analyze large primary care datasets for COVID-19 pharmacoepidemiology research. The authors propose a consensus statement to ensure transparency, rigor, and consistency in defining cases, exposures, and outcomes, enabling more reliable comparisons and meta-analyses of findings.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for rapid, large-scale research to inform patient care and policy. Primary care electronic health records (EHRs) represent a vast, rich source of data for this purpose, having been previously used for pandemic surveillance and intervention studies. However, a significant challenge emerged: the lack of a standardized approach to conducting and reporting pharmacoepidemiological research using these big data resources for COVID-19. This absence of consensus leads to inconsistencies in how cases are defined, how drug exposures are characterized, and how outcomes are measured, hindering the ability to compare, validate, and synthesize findings across different studies. The paper aims to bridge this gap by providing a framework for consistent and transparent research practices.

Methods

The paper's core contribution is a consensus statement developed by a consortium of UK big data experts, epidemiologists, researchers, statisticians, and frontline clinicians. The methodology involved:

  • Consortium Development: A group of experts from various UK universities and clinical backgrounds collaborated.
  • Iterative Refinement: Weekly meetings were held to refine ideas and reach agreement on key items for inclusion in the statement.
  • Focus on Standardization: The consensus focuses on establishing common definitions and reporting guidelines for critical elements of pharmacoepidemiological studies, including:
  • Protocols: Encouraging the public availability of research protocols and analysis plans before data analysis to promote transparency and reduce duplication.
  • COVID-19 Infection Definitions: Proposing specific definitions for "Confirmed COVID-19 Case," "Suspected COVID-19 Case," and "Suspected COVID-19 With a Negative Test," acknowledging the limitations of operationalizing these in EHRs.
  • Drug Exposures: Outlining principles for reporting drug names, classes, chemical names, combination preparations, exposure time, prescription details, and dosage.
  • Confounding Variables: Recommending a list of variables to consider as confounders, with flexibility in categorization based on data availability.
  • Stratification Variables: Suggesting stratification by age, sex, ethnic group, and domicile to examine disproportionate impacts.
  • Outcome Reporting: Endorsing the use of established reporting guidelines (STROBE, RECORD, TRIPOD) and providing a table of recommended short-term and long-term outcomes for both primary care and ICU settings.
  • Analytical Methods: Emphasizing transparency in reporting analytical methods, including hypothesis, sample size, descriptive characteristics, model assumptions, missing data management, and sensitivity analyses.

Datasets

The paper discusses the use of several large primary care databases in the UK by the COVID-19 consortium:

  • QResearch: Established in 2003, with over 10.6 million patient records (expanding to 21 million).
  • RCGP Research & Surveillance Network Center: Established in 1957, with 5 million patient records.
  • Clinical Practice Research Datalink (CPRD): Established in 1989, with 14 million patient records.
  • UK Biobank: Established in 2006, with partial cohort coverage of 0.5 million records.

These databases are variously linked to public health, secondary care, and vital status records. The paper notes that there is likely overlap between these datasets, with approximately 20% of patients appearing in at least two datasets.

Key Insights and Ideas

  • The Power of Primary Care Big Data: The paper underscores the immense potential of routinely collected primary care data for rapid, large-scale research during a pandemic, offering insights into disease surveillance, risk factors, and intervention effects.
  • The Critical Need for Standardization: The lack of a consensus on definitions and reporting methods is a major bottleneck for reliable COVID-19 pharmacoepidemiological research. This statement aims to rectify that.
  • Transparency as a Cornerstone: Making research protocols and analysis plans publicly available is crucial for transparency, scrutiny, and reducing false positives.
  • Nuance in COVID-19 Definitions: Operationalizing COVID-19 definitions in EHRs is challenging due to limitations in coding clinical symptoms and the evolving nature of testing. The proposed definitions aim to strike a balance between rigor and practicality.
  • Comprehensive Drug Exposure Reporting: Simply identifying drug use is insufficient; detailed reporting on drug names, classes, exposure duration, and dosage is essential for accurate pharmacoepidemiological analysis.
  • The Importance of Confounders and Stratification: Acknowledging and appropriately handling confounding variables and stratifying analyses by key demographic factors are vital for understanding true treatment effects and identifying disparities.
  • Balancing Speed and Rigor: The urgency of the pandemic necessitates rapid research, but this must be balanced with rigor and external validation of emergent findings.

Key Results

This paper does not present quantitative results from specific studies. Instead, its key "results" are the consensus recommendations themselves, which include:

  • Standardized Definitions: Agreed-upon definitions for confirmed, suspected, and negative COVID-19 cases.
  • Reporting Guidelines: Detailed principles for reporting drug exposures, confounding variables, stratification variables, and outcomes.
  • Methodological Recommendations: Guidance on analytical methods to ensure transparency and rigor.
  • Endorsement of Reporting Checklists: Encouragement to use STROBE, RECORD, and TRIPOD checklists.
  • Table of Recommended Outcomes: A structured list of short-term and long-term outcomes to report for COVID-19 studies.

Limitations and Future Work

Limitations:

  • Data Limitations: The paper acknowledges inherent limitations of EHR data, including potential inconsistencies in coding accuracy, completeness, and the delay in outcome data reaching GP records.
  • COVID-19 Coding Uptake: The uptake of newly introduced clinical codes specific to COVID-19 may not be universal.
  • Asymptomatic Cases: The data may under-represent asymptomatic or less severe cases, potentially over-representing disease severity and contributing factors.
  • Testing Limitations: UK testing for COVID-19 has been limited, and the sensitivity and specificity of tests are still being validated.
  • Evolving Definitions: COVID-19 definitions are evolving, and the proposed definitions are based on the understanding as of March 13, 2020, with acknowledgments of their limitations.
  • Causal Inference: Making causal inferences from non-randomized observational data requires careful interpretation and appropriate analyses.

Future Work:

  • Updating Drug Lists: The list of UK-prescribed drugs requiring urgent characterization will be updated on the QResearch website as more data become available.
  • Expanding Stratification Variables: Further stratification variables should be considered as more data are published.
  • Understanding Sequelae: As the short- and long-term sequelae of COVID-19 become better understood, additional outcomes in primary care will need to be added to the reporting list.
  • Exploratory Analyses: While the consensus focuses on inferential methods, the paper acknowledges the increasing value of exploratory or data mining analyses.
  • Global Application: The consortium's work could inform similar approaches worldwide, establishing linked primary care records in individual countries and enabling cross-country comparisons.

Summary updated

Abstract

The use of big data containing millions of primary care medical records provides an opportunity for rapid research to help inform patient care and policy decisions during the first and subsequent waves of the coronavirus disease 2019 (COVID-19) pandemic. Routinely collected primary care data have previously been used for national pandemic surveillance, quantifying associations between exposures and outcomes, identifying high risk populations, and examining the effects of interventions at scale, but there is no consensus on how to effectively conduct or report these data for COVID-19 research. A COVID-19 primary care database consortium was established in April 2020 and its researchers have ongoing COVID-19 projects in overlapping data sets with over 40 million primary care records in the United Kingdom that are variously linked to public health, secondary care, and vital status records. This consensus agreement is aimed at facilitating transparency and rigor in methodological approaches, and consistency in defining and reporting cases, exposures, confounders, stratification variables, and outcomes in relation to the pharmacoepidemiology of COVID-19. This will facilitate comparison, validation, and meta-analyses of research during and after the pandemic.
BibTeX
@article{dambhamiller2021use,
  author = {Hajira Dambha‐Miller and Simon J. Griffin and Duncan Young and Peter Watkinson and Pui San Tan and Ashley Kieran Clift and Rupert Payne and Carol Coupland and Jemma C. Hopewell and Jonathan Mant and Richard M. Martin and Julia Hippisley–Cox},
  title = {{The Use of Primary Care Big Data in Understanding the Pharmacoepidemiology of COVID-19: A Consensus Statement From the COVID-19 Primary Care Database Consortium}},
  journal = {The Annals of Family Medicine},
  year = {2021},
  doi = {10.1370/afm.2658},
}
Figure