Duke-Margolis white paper outlines best practices for EHR data curation in regulatory real-world evidence

A new white paper from the Duke-Margolis Real-World Evidence Collaborative analyses how electronic health record (EHR) data are curated for regulatory use, outlining conceptual clarifications, operational best practices, and governance considerations for AI.

The Duke-Margolis Institute for Health Policy’s Real-World Evidence (RWE) Collaborative has published a white paper titled ‘Data Curation Best Practices and Innovations for Real-World Evidence Generated Using Electronic Health Record-Sourced Data’. The document summarizes a 2025 workstream on “Operationalizing Relevance, Reliability, and Quality During Electronic Health Record Sourced Data Curation,” building on a 2024 initiative focused on data accrual. Through monthly meetings and a dedicated workshop, participants examined how EHR-sourced real-world data (RWD) can be curated to preserve clinical meaning and meet regulatory expectations.

The paper frames its discussion within the US FDA’s three-stage EHR data life cycle: accrual, curation, and transformation. FDA guidance defines data curation as “the processing of source data through the application of standards for exchange, integration, sharing, and retrieval,” encompassing activities such as cleaning, error correction, terminology standardization, and system integration. The authors note that while this definition is conceptually sound, curation in practice represents the initial step in a broader, more complex data management process involving multiple data holders and intermediaries.

A key regulatory principle highlighted in the paper is preservation of clinical meaning. The authors write that:

“Data curation processes should not alter nor obscure the clinical meaning of the source data and preserve important contextual information.”

The white paper emphasizes that maintaining this standard becomes more challenging when data are mapped and harmonized across platforms, particularly when structured and unstructured elements from different systems are combined to achieve interoperability and analytic usability.

The authors observe limited consensus on how to operationalize regulatory expectations across the three life cycle stages. They state that:

“Unstandardized or unclear data curation practices can limit the regulatory utility as well as relevance and reliability of EHR-sourced data.”

Variability in documentation practices, coding conventions, and local workflows may introduce inconsistencies that affect regulators’ ability to assess exposures, outcomes, and covariates.

The paper questions whether curation should be seen as a separate, intermediate step. It notes that the move from source data to an analytic dataset “is neither straightforward nor neatly compartmentalized.” Tasks such as cleaning, harmonizing data, and creating proxy variables often take place in stages and may be repeated over time. The authors therefore suggest that:

“It may be useful to view curation as an iterative, cross-cutting process rather than as a discrete, monolithic phase.”

The paper also discusses interoperability policy in this context. While the 21st Century Cures Act required the use of application programming interfaces to improve data exchange, it did not require common coding systems or standard definitions for data elements. Similarly, although HL7 FHIR may improve exchange, it “primarily supports data exchange” and does not by itself assess data quality or address differences in how data are recorded locally.

To support regulatory confidence, the authors emphasize the need for structured and auditable processes. They highlight predefined curation procedures, audit trails, and clear documentation of data provenance as ways to improve transparency. The recommendations include a call to:

“Implement standardized, auditable, and harmonized standard operating procedures (SOPs) integrated with FHIR to promote consistent validation and traceability procedures across institutions.”

The paper further examines the expanding role of AI in EHR data curation. AI tools are increasingly used to convert unstructured clinical notes into structured variables and to identify inconsistencies or harmonize coding within structured datasets. However, the authors note that “its uptake across the healthcare ecosystem has drastically outpaced the development and broad adoption of data standards, validation frameworks, transparency guidelines, and general governance.” They add that:

“Granular guidance specifically geared toward AI in EHR-sourced data curation remains underdeveloped.”

Concerns include limited transparency in model decision-making, reduced visibility into how inputs are processed and outputs generated, and the potential for system-level bias when training data are disproportionately derived from a single EHR system.

Human oversight is presented as essential. The white paper states that “AI tools should not autonomously modify EHR data but instead be designed to support human decision-making,” and emphasizes that:

“Humans remain central and accountable to the AI-assisted decision-making process.”

Recommended practices include documenting training data characteristics, reporting validation metrics such as agreement between human and AI workflows, and following principles of transparency and interpretability in AI-enabled curation.