Case 1: Researcher who creates a population-based study

Researchers who generate data in a cohort, case control or clinical study represent a large portion of “data generators”, with laboratory-based researchers representing a second group (see Case 2). Researchers who collect data have traditionally acted as gatekeepers to data use-which has led to criticism. Those who collect data spend years developing ideas, collecting pilot data, writing grants, hiring and managing personnel, writing IRB proposals, responding to critiques, creating a database, cleaning and creating a data dictionary before analyzing the data themselves.

High quality data generation is an area requiring considerable expertise. Longitudinal data are the gold standard in human subject research and can require decades of effort. This can lead to an imbalance in the amount of work hours needed to generate data versus work hours needed to analyze shared data. Traditional academic systems prioritize publishing research papers rather than sharing data which compounds this problem. Without safeguards for authorship consideration, the present scenario creates negative incentives in which those generating data are reluctant to share, fearing lack of recognition. Further, researchers using extant data analysis for manuscript writing may be years removed from the data collection process. They may not understand or appreciate the strengths and weaknesses of the dataset or may make incorrect assumptions about how the data were generated. Authorship criteria require substantial contributions to the conception, design, execution, or interpretation of the research. Generating data by itself often meets these criteria.

Case 2: Data sharing from a laboratory perspective

A laboratory may generate data from analysis of biosamples without conducting sample collection. Such researchers often have unique expertise and insights on the assay and its interpretation. Assays require processing and researchers may sometimes submit raw data to meet data sharing requirements. This can restrict its use to only researchers who can re-process results, while excluding researchers without such expertise. In many instances, considerable effort has been expended by laboratory faculty and staff in developing methods specific to the sample analysis, as well as ensuring that methods are executed with proper quality control. Without proper annotation and safeguards, data users may not recognize the laboratory contribution to the deposited data. If team science is to be encouraged, appropriate credit should be provided to laboratory investigators. Their contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data for the work merit authorship² and often meets the standard for authorship.

Case 3: Data sharing from a data user perspective

Data generators may not have advanced skills in data science. New structures of exposome data and larger datasets will require the development of relevant data science methods. We have seen this in the recent interest in environmental mixtures and the surge of statistical methods developed in response. Exposome data may be assessed from heterogeneous biomarkers among many different scenarios. Different methods are needed for mixtures data that are observed cross-sectionally versus data assessed longitudinally as repeated measures of exposure or grouped by life stage or chemical class. In addition, different methods are needed to address different research questions.^3,4 The vast majority of statistical and data science methods are developed by data scientists to match the data structure and questions that can naturally be answered from the study to which they have access. In part, this is because demonstrating methods using real data is often required for publication. Sharing data is a key process for overcoming this roadblock to develop methods for emerging exposome data structures and to answer new questions within existing data sources. As new data structures become available, it is important to share these data so that appropriate methods can be developed to properly analyze data.

Data sharing is essential for benchmarking and comparing exposome analysis methods which informs researchers about the relative abilities and operating characteristics of different methods, including which methods are most appropriate for their analyses. A recent example is the 2021 Exposome Data Challenge Event.⁵ This event gathered a diverse scientific audience, including epidemiologists, biostatisticians and computational scientists applying a variety of data science tools to address different research questions and inferential goals related to exposomics, allowing for comparisons of methods. As a result, teams illustrated applications of methods on a common dataset and shared code. Other examples from environmental mixtures research include a National Institutes of Health (NIH) workshop that compared methods on real and simulated datasets.⁶ The workshop include simulation studies that quantitatively compared performance in simulation studies⁷ and papers that qualitatively compared methods on select datasets.⁸ Sharing data allowed for benchmarking for new structures, addressing new research questions, and avoiding biases caused by overusing a limited number of data sources to validate model performance. Innovative means of data sharing include federated data analysis, which is performed on multiple separated datasets without data exchange, maintaining the safeguards of an institution’s firewall. Only parameters of the analysis method are exchanged so that sensitive information remains anonymous. Such methods reduce the need for transferring data, reducing risks of compromising personal information.

Case 4: Data sharing internationally

International collaborations are a special case of data sharing. While the possibilities for collaboration are limitless, we present Europe as a case study. The European Health Data Space (EHDS) is a recent European Union (EU) regulatory initiative to support the use of health data. Updated guidelines regulate data sharing beyond EU countries and apply to data collected and shared with a NIH funded study. The Trans-NIH BioMedical Informatics Coordinating Committee maintains lists of domain-specific data repositories in the US, while similar data repositories exist in the EU (eg, Zenodo, etc,).

Beyond EHDS, the European General Data Protection Rule (GDPR), requires protection be ensured when data are transferred to third countries (eg, US). Regardless of the data sharing tool used, transferred data must receive an equivalent level of protection as it would in the European Union. The EU Commission and the USA recently agreed on a Trans-Atlantic Data Privacy Framework (TADPF) to ensure that US surveillance activities are necessary and proportionate in pursuit of national security objectives, by creating an independent complaint mechanism for European citizens.⁹

Case 5: Sharing real-world data

We conclude with the concept of reuse of administrative data i.e., not collected for research- another special case of data sharing. With the proliferation of Information and Computing Technology systems in health (eg, Electronic Health Records), the volume of health data has dramatically increased. Such datasets will be used for purposes different than those for which they were collected, such as epidemiological studies, policy making, etc, These data are called “Real World Data” (RWD) and the conclusions inferred upon them are called “Real World Evidence” (RWE).^10-12 Their importance is widely recognized despite methodological challenges regarding their quality and verifiability. However, using RWD is not trivial as they are often “sensitive” personal data, thus accompanied with legal, ethical, administrative issues. In technical terms, “interoperability” is also a crucial issue, as data exchange or federated analysis would require the use of a Common Data Model (CDM).^13,14

There are several initiatives supporting the development of observational studies with RWD, including developing a “research network” of Data Partners (DPs) with data available in a specific CDM compatible format. Instead of collecting the data from DPs in a central analysis node, only the non-sensitive and non-personal aggregated results are collected. This prevents moving sensitive data out of the original host/DP reducing legal challenges. Only the query and its aggregated results are transferred enabling observational studies in a federated fashion without sharing raw patient data.

Along these lines, Observational Health Data Sciences and Informatics (OHDSI) should be commended as a global initiative aiming to facilitate the analysis of RWD via observational studies.¹⁵ OHDSI developed a set of open-source tools based on the so-called “OMOP Common Data Model” (OMOP-CDM). The European Health Data & Evidence Network (EHDEN) acts as the European face of OHDSI¹⁶ setting up a community of data partners across Europe upon OMOP-CDM, currently including more than 180 organizations. As a whole, the OHDSI ecosystem has proven that technical means can go a long way in terms of overcoming legal and regulatory barriers in an international scale.^20-22

Even though technical standards may facilitate the syntactic interoperability between data silos (eg, hospitals), the issue of semantic interoperability still remains. To this end, widely accepted reference terminologies are used (eg, the Unified Medical Language System—UMLS). Developing and maintaining such terminologies/ontologies and aligning them is a costly and error prone process. Data quality is also a significant issue as research and observational studies are secondary use cases for RWD and not the priority of the original data creators. Contextual factors should be considered when RWD are interpreted.

Recommendation 1

Authorship should be a consideration for data generators when sharing is mandated.^17-19,23 This does not mean automatic authorship but a system that facilitates and requires communication between data generators and data users. This system should distinguish between expert (longitudinal study, novel or complex lab assay) and RWD data generation. Citing data that has been deposited is also recommended by the Joint Declaration of Data Citation Principals.²⁰

Recommendation 2

The costs of data sharing should be shared. Implementation of fees to use data would help cover some costs of data sharing. Even with such fees, significant infrastructure investment is needed to support data harmonization, ontologies and search engines that allow researchers to find datasets. Data sharing fees may not be sufficient. Training programs in these methods are needed if we are to reach a critical mass of data generators and data users.

Recommendation 3

Ethical issues in sharing data need guidance from the Federal government to ensure privacy. At present, there are few guidelines to assist IRBs in determining whether data can or should be shared and ethical issues may be highly contextual requiring development of an expert review board to evaluate cases as they arise. Adding “data donation” routinely to consent forms as an opt out might accelerate data sharing.

Recommendation 4

Infrastructure is needed to facilitate federated data analysis as some data cannot be shared in a centralized database. Finally, data repositories should facilitate advanced computational approaches that enable full processing of data while still keeping data protected. Methods beyond deidentification need to be developed in scale. For example, homomorphic encryption allows data processing while keeping data encrypted, but its use is computationally intensive.