Sharing participant-level trial data
Participant-level trial data reach a secondary user either on request or through a repository or platform that holds them, and either way the data set has to be documented well enough to be used and protected so that no participant can be identified. A writer's part lies in that documentation, in consent and registry wording that allows later sharing, and in documents that need little redaction. Technical, motivational, economic, political, legal and ethical obstacles decide whether data are shared at all, some kinds of data and public health emergencies strain the usual mechanisms, and the same data are increasingly re-used inside the companies that generated them.
The two sharing mechanisms
Clinical trial data are shared by two principal mechanisms. In the first, investigators share trial data on their own terms in response to individual requests. Someone seeks out a team that generated trial data on a subject of interest and asks to use them, perhaps as part of a larger meta-analysis, to validate the original findings or to pursue a secondary question the original team did not address. Each request is decided on its own, and an approved request leads to a direct transfer between the teams. Sharing of this kind is less predictable, because it is never clear whether or when a request will come.
In the second mechanism, investigators deposit the data in a repository, an archive whose terms of access are set by the organisation that maintains it. For deposit to be effective, the data have to go into a publicly accessible repository, such as DRYAD, where they can be used under the rules that govern access. Deposit needs a standardised approach and is predictable in practice, because the data have to be prepared for sharing whether or not anyone ever asks for them. The decision to share is made in advance, and the data are transferred after the study is completed and all data management and preparation issues are resolved. The deposit holds the data files together with their metadata and documentation.
Why participant data are shared
Data underlying clinical research had historically stayed in the custody of the data generator, in most cases a pharmaceutical or academic sponsor, and access was highly limited. Sharing was not new: academic institutions, the pharmaceutical industry, regulators, pricing agencies, patient groups, investigative journalists and the media had encouraged it for some time. Sharing benefits patients by pointing to new research questions that can lead to new discoveries, and it lets trial results enter meta-analyses that raise standards of evidence and lets published results be confirmed, which reduces bias. It also honours the generosity of participants by making their data more useful and so their contribution more valuable. The main goal of sharing participant data is to let other researchers repurpose them for secondary uses. Access can let a study be independently replicated, prevent duplicative studies, give the basis for generating or testing new hypotheses and generally advance medicine and biology.
By June 2019 sharing de-identified or anonymised participant data was described as rapidly becoming the norm. The full effect of the requirements for sharing, and how they fit with other legal provisions, especially the protection of participants' personal information and of sponsors' commercially confidential information, had still to be resolved.
Sharing can be justified on scientific, economic and ethical grounds. In 2019 large repositories of participant data and technologies able to analyse large data sets were becoming available, and national legal interpretations of the laws on sharing were expected to be resolved and harmonised. Sharing de-identified or anonymised participant data was held to be here to stay and to keep developing and expanding. The benefits to patients have to be kept in view while the privacy of the participants who contribute their data is respected.
By June 2020, as seen from the sponsor side, sharing individual participant-level data had become an almost routine extension of the clinical study life cycle and increasingly a vital element of leveraging real-world data. Key research funders, journal editors, pharmaceutical trade associations, regulators, ethics committees and government sponsors of research expected responsible sharing of appropriately consented and de-identified participant data and associated documents, and patients increasingly supported it. Journals' requirement to make a data sharing plan public as a condition of publication, in line with the International Committee of Medical Journal Editors and the PLoS One journals, gave every data generator a further incentive. Harnessing the statistical power of large data sets was envisaged to let the broader scientific community embrace the big data revolution, including machine learning and artificial intelligence.
Industry commitments to share
On 1 January 2014 the European and US industry associations released joint principles for responsible clinical trial data sharing. Under them, researchers can submit proposals for access to patient-level data, protocols and clinical study reports for new medicines approved in the US and the EU after that date. The international federation adopted similar commitments on 15 January 2018. Many sponsors, academic institutions and companies outside the associations among them, updated their disclosure policies to include consideration of sharing with qualified external parties. Most pharmaceutical sponsors set the terms in company policy, for example sharing only for products approved in both the US and the EU or for completed trials, some of which have many years of follow-up before they count as complete.
By 2016 most of the largest pharmaceutical companies had publicly committed to responsible data sharing, all members of the European industry federation but few non-members. In August 2016 a search of the public global websites of the top 50 pharmaceutical companies by 2014 global sales looked for statements on data sharing and on compliance with the federation's principles. Of the 50, 27 were members of the federation, three of them affiliates. A data sharing policy was found on every member's website and on 4 of the 23 non-members'. An explicit reference to the federation's principles was found on 22 of 27 member and 1 of 23 non-member websites, and references to all five principles on 15 member websites and 1 non-member website. By commitment, members and non-members respectively referred to sharing trial data with researchers in 25 of 27 and 2 of 23 cases, making clinical study report synopses publicly available in 23 and 1, making results available to trial participants in 24 and 1, publicly certifying adoption of the federation's commitments in 26 and 1, and committing to publication of trial data in 26 and 3. A cross-sectional study of data sharing and results reporting among large pharmaceutical companies, published in 2019, found that some companies had made noteworthy commitments to share participant-level trial data and others were willing to improve their policies, but many still had substantial room for improvement.
Data sharing platforms
Data sharing platforms are an alternative way for trial sponsors to share participant data. Several joint initiatives set up repositories to which subscribing sponsors hand the documents and data sets of selected trials. Each platform sets what a request must contain, the format the data sets arrive in and the working site where secondary analysis is done. Sponsors pay a fee to take part in some of them, which handle most of the assessment and processing of requests. The platforms help sponsors meet the ethical obligation to share de-identified or anonymised participant data. Platforms in 2019 included the ClinicalStudyDataRequest consortium, the YODA Project, Vivli, Project Data Sphere, which charges no fees, and DataCelerate, beside national or institutional platforms such as the US National Institutes of Health's and disease-specific ones such as the Alzheimer's Disease Neuroimaging Initiative. It was too early to draw definitive conclusions about the platforms' effectiveness. Membership was relatively low and its costs high, the platforms were not interoperable, and the participant data available from sponsor-selected trials were fragmented, so identifying, obtaining and evaluating all the relevant documents remained hard for an external requester. Vivli is one of several platforms offering an alternative to the perhaps more widely known ClinicalStudyDataRequest. On 13 May 2019 it hosted a meeting in Tokyo on the implications of the General Data Protection Regulation for data disclosure and data sharing one year after the regulation took effect.
ClinicalStudyDataRequest.com was started by GlaxoSmithKline in May 2013 as a consortium of clinical study sponsors offering patient-level data through a researcher-friendly platform with independent review of proposals and protection of patient privacy. The Yale University Open Data Access (YODA) Project, from October 2014, offers responsible sharing of clinical research data, open science and research transparency, and supports research aimed at improving patients' health and informing science and public health. Vivli, started by the Multi-Regional Clinical Trials Center of Brigham and Women's Hospital and Harvard in November 2016, has an independent data repository, a search engine and a secure research environment, in which users search listed studies, request data sets, aggregate data and share data of their own. Such platforms, from companies, universities and non-government organisations, let researchers request individual patient data for secondary, independent analyses.
Ad hoc, fit-for-purpose sharing had been occurring in the pharmaceutical industry for some time, but large-scale, coordinated sharing gathered momentum after the trade associations' principles for responsible data sharing of 2014. Early adopter companies built portals to accept requests, independent review panels to decide on access and secure research environments, and these efforts led to ClinicalStudyDataRequest.com and the Yale Open Data Access Project. Duke Clinical Research Institute's Supporting Open Access for Researchers platform is a similar partnership with sponsors and others. These portals essentially follow the same major steps, study listing, proposal review, approval by an independent review panel, a data sharing agreement and research conduct, although some aspects of the Duke platform emphasise curation and data harmonisation. Most pharmaceutical and biotech sponsors, though, share through stand-alone portals, from proposal gateways to online forms or email, with company-specific review panels. Pre-competitive sharing among sponsors is a separate arena, meant to let them collaborate towards more efficient and effective clinical development. The IQ consortium shares largely technical preclinical and early clinical data to identify new science, technology and regulatory engagement pathways, and DataCelerate supports sharing among TransCelerate and BioCelerate members. The proliferation of platforms left mechanisms with limited interoperability that are inefficient and hard for a researcher to navigate, and efforts towards cross-platform access were under way.
Among its COVID-19-related initiatives, TransCelerate Biopharma, a group of member biotechnology and pharmaceutical companies working to standardise key clinical trial activities, released DataCelerate, a data-sharing platform for COVID-19 research.
Non-profits entered both as funders and as infrastructure, to make access easier, lower the threshold for entry and make the process more independent. The Wellcome Trust enforces data sharing requirements on the research it funds and has supported ClinicalStudyDataRequest.com and Vivli, primarily by supporting the management of their review panels. Project Data Sphere, created by the CEO Roundtable on Cancer's Life Sciences Consortium, began by sharing downloadable comparator-arm data from oncology studies and can also host data in a secure environment or through Vivli. Vivli seeks to serve as a neutral broker between data providers, data accessors and the wider sharing community, and its platform seeks to span all disease areas. It intended to add disease-specific communities and support for de-identification that would lower the barrier to entry for smaller and academic data providers. The efficiency and impact of sharing can be maximised when data sets are curated and harmonised, so data sharing is increasingly an important part of large public-private partnerships, which are coordinated by governmental and industry or academic bodies, funded jointly and generally focused on specific public health priorities. They include the Innovative Medicines Initiative between the EU and the European industry federation, more and more of whose projects aggregate data on specific diseases. They also include the Accelerating Medicines Partnerships, funded through the Foundation for the National Institutes of Health with the US industry association's support to accelerate research in a range of disease areas and support biomarker development, and the Critical Path Institute, a non-profit partnership involving the FDA that aggregates data in pre-competitive consortia to create new data, measurement and methods standards.
Trade associations and other organisations have also led data sharing initiatives. The European, international and US industry federations, EFPIA, IFPMA and PhRMA, encourage their members to share clinical data by disclosing study results, giving access to individual participant data and making clinical study reports available for secondary analyses, in a way that protects patient privacy and respects commercial confidentiality. Vivli, clinicalstudydatarequest.com and the Yale Open Data Access Project are independent platforms that give external investigators access to clinical trial data from multiple sponsors. The World Health Organization endorses the timely release of clinical trial results and supports platforms that facilitate data sharing. Together these initiatives aim to maximise the utility of clinical data, foster innovation and improve health outcomes globally.
The request process
Patient-level data sharing is the process by which data providers accept requests from academic and non-academic researchers for access to data and supporting documentation from formal clinical trials, and its common steps vary with the provider or platform. A request starts from knowing what data exist, so providers need to make public their data sharing policies and which studies are available. The trade association principles require sponsors to make their data sharing policies public and to provide a way of receiving research proposals. ClinicalTrials.gov had added an individual participant data sharing plan section that each sponsor must complete at registration. By June 2020 most pharmaceutical and biotechnology sponsors had published their policies and processes, and most pharmaceutical data providers had joined consortia or built stand-alone portals listing the studies available. Data generators of every kind increasingly published their sharing plans at registration and publication. Together these measures had dramatically increased the discoverability of studies available for data sharing. Providers also make sure in advance that the data can be shared: that they are de-identifiable, consented and contractually permitted to be shared.
A request is then a research proposal, usually from a researcher on behalf of a broader team that includes a statistician or another qualified data analyst, such as a health economist. It names the studies wanted and any other data to be combined with them, and it often carries a data management plan, a statistical analysis plan, a publication plan and a conflicts of interest statement. The data provider usually checks the proposal first for completeness and against its own policy, the study's informed consent and other legal bases for sharing, such as the General Data Protection Regulation. An independent review panel then assesses scientific merit and other aspects, such as conflicts of interest and the researchers' qualifications. The panel acts either as the primary reviewer or as an appeal route for requests the sponsor rejects, and its role and independence vary. An approved requester signs a data sharing agreement that sets the access conditions and licences, including commitments to protect participant privacy and the provider's confidential information. The requester then receives anonymised data and de-identified documents, protected to minimise the risk of re-identifying participants and of releasing company confidential information. Access is usually given in a secure cloud-based research environment and for a defined period, usually one to two years, with extensions possible. The researcher conducts the project to the proposal's analysis plan, asks for an extension or an approved amendment where needed, and tells the provider or the panel of any new safety signal, intellectual property or publication. Access closes as the agreement specifies, and the manuscript is submitted. Not every provider uses every step, and some add one, such as an appeal against refusal. "Data provider" covers academic and industry sponsors, other data generators such as non-profit research bodies, and data stewards who manage data they acquired but did not generate.
A request to the clinicalstudydatarequest.com consortium portal is a proposal of what the requester means to do with the data, in effect a protocol for a study, such as a meta-analysis, although no data are being collected. The portal publishes metrics on how many applications it received, approved and refused, and why, which give some idea of the process.
A request to a company for additional analyses of its data, made through EFPIA in Europe or PhRMA in the United States, has to include a publication plan for the proposed re-analysis.
Documentation for secondary users and the writer's part
Both sharing mechanisms were expected to involve medical writers more and more, particularly for industry-funded trials and as sharing became standardised. Writers can play a critical role in several steps of sharing, but their chief responsibility is likely to be preparing the metadata and documentation that secondary users of shared data need. Structural metadata cover the design and specification of the data structures, and descriptive metadata cover information about the data content. A secondary user needs both to find their way around the structures and the content, which improves the usability of the data.
The data set itself has to be clean and well annotated, with supporting documentation that allows secondary analysis. It has to be clearly organised, follow standard logic and coding formats, and define every variable without ambiguity. Without strict adherence to best practice in preparing and documenting data, sharing will be largely limited.
Standards for data definition also have to be adopted across trials, so that data can be compared and pooled for summary analysis. Several trials of acute myocardial infarction that each define the endpoint differently limit any attempt to summarise or interpret the data across them. Policies and procedures for preparing data for deposit varied among funders, and few were making a concerted effort to adopt uniform ones. If every funder adopts different data definitions and documentation requirements, sharing may become an inefficient use of resources, time and energy. For sharing to succeed, policies and procedures have to be standardised. Medical writers can identify the most appropriate and effective policies, help establish the standards for data definition and documentation, advocate that funders standardise, and lead the development of an efficient and feasible system for sharing. Those efforts will ensure that the research community derives the full benefit of the enormous resources devoted to clinical trial research and will help build patient trust in the research process. The data and information that trials generate have to be available to the scientific and clinical communities so that treatment decisions rest on all that is known.
Writers can enable later sharing by considering data sharing when drafting protocols and informed consent forms. Telling participants the sponsor's sharing plans, the possible uses of the data and the residual privacy risk of sharing de-identified data can substantially enable later sharing, and such language can help clarify the legal basis for sharing under evolving privacy law and streamline ethics committee approval. Clear data sharing plans on trial registers and in publications improve discoverability. Lean writing can minimise the need to redact while still producing documents of high clinical utility, and documents that are easier to protect can be more useful to researchers and make sharing more efficient. Other enablers of efficient sharing are broader use of common data standards and releasing ahead of time the structure and contents of the data to be shared, the data dictionary or metadata, alongside the listed title, so that researchers can plan and understand what they are getting. Data sharing made immense progress in the five years to 2020, but more could be done, especially given emerging disease challenges that will need data-driven solutions. Medical writers are well placed to be key contributors to that progress, including by describing the legal basis for sharing clearly and writing with privacy protection in mind.
Sharing and its consequences concern not only medical writers, who collate and describe the trial data, but also the data managers and statisticians who collect, collate and process participant data. The technical and statistical challenges of accessing data for reanalysis are not trivial, and converting collected data into sets an external researcher can analyse needs specific skills and techniques. Statisticians should move from guarding data to facilitating access to it.
Technical questions in preparing data
Preparing data for sharing raises technical questions that are likely best handled by medical writers together with the other investigators and the data analysts. They are potential challenges facing the field.
What counts as the data set. Far more data are collected in a trial than are reported in any single journal article. The limited guidance available suggests that the data set is the aggregated collection of patient observations, socio-demographic and clinical information included, used to produce the summary statistical findings in the main report of the project, whether or not previously published. On that definition it should include all pre-specified and intentionally collected primary and secondary outcomes as well as safety endpoints, in agreement with the trial registration and results reporting requirements of ClinicalTrials.gov.
How far de-identification has to go, and what it costs. To protect patient confidentiality, all direct and indirect patient identifiers must be removed from the data set. Experts disagreed on how complex the task is, and presentations at an Institute of Medicine workshop suggested it could take anything from an hour to an afternoon to weeks. Confidentiality is, however, of greater concern for small trials in rare diseases than for large multi-centre trials of common conditions, where identifying a patient is extremely difficult.
Whether statistical output has to be revised for transparency is a question at the level of the data set, where much work on how to disclose was under way. Data sets have to be anonymised before they can be disclosed, and recommendations for anonymising them can be found, for example, on the TransCelerate website and on that of a data management group. EMA Policy 0070 itself refers to material that is freely downloadable on the web. In one practitioner's view, medical writers can apply anonymisation techniques themselves, but statistical outputs need special software, so at that level the work depends on data management experts and the software they use.
Who owns the data. Ownership has to be resolved before deposit. A funder may settle it in advance by requiring deposit in a repository and transferring any ownership held by the funder or the research team to the public. Because most clinical research data are generated through collaborations between several researchers, nearly all of whom are paid for their effort, it may not be possible to determine who owns them, and some contend that the patients are the rightful owners. The best guidance available came from recommendations on publishing raw data in journals: copyright should be transferred to the publisher for data sets published as supplementary material, the supporting data should be separated from the article itself, and transfer of copyright for the data is not required as a condition of publication. Individual repositories are likely to set their own policies.
Whether the participants agreed to it. Data sharing plans were rarely discussed with patients during informed consent. De-identification may remove the need for consent before sharing, but ethics committees and institutional review boards should encourage researchers to discuss data sharing plans when obtaining consent, along with any safeguards that will protect privacy. Everyone's interests would be best served if patients explicitly consented to the sharing of their de-identified research data.
Obstacles to sharing
Barriers to data transparency can be technical, motivational, economic, political, legal and ethical. Sponsors often resist sharing on the ground that having paid a great deal for their trials makes them the owners of the data, but that is only one of the obstacles.
Technical. Data may not be collected at all, and around the world they are collected in different ways. In the COVID-19 registries of April 2020, countries' differing death and hospitalisation rates could simply have reflected how each collected its data. Disease surveillance systems did not meet the standards set by the International Health Regulations in 2005. Data that were collected may not be preserved, because healthcare providers gather most of them for short-term purposes and preservation is usually not a priority. Some data were collected with software built on technology that no longer exists: at one company a system had not worked for about twenty years, and where its data had gone was not known. Data collected in different languages, even within one country such as Switzerland with its four official languages, are very difficult to put together. Format is another barrier: paper collection persists in some countries and some therapeutic areas and has to be keyed into an electronic system before anything can be analysed. Software systems may not talk to each other or integrate, and data sets can arrive in formats too variable to pool. Metadata standards are needed, and they were being established but were far from widely implemented around the world.
Motivational. In one practitioner's view this is the highest-order obstacle. Sharing data needs people to run the process, and so time and resources, which most organisations, pharmaceutical companies included, had not planned for. Incentives are needed for people to give sharing priority, and one intent of the journal editors' request for raw data to be shared with the final publication is to provide one. An academic who wants to publish in a high-tier journal must disclose the individual patient data so that anyone can replicate the work. The capability is unevenly spread. To that practitioner's knowledge only a few universities had the technology, and among pharmaceutical companies it was well established in the larger ones and not in others. In medical devices some large companies had set up a department for these documents while others had nothing, which the newer device rules could make a serious problem. A researcher who has invested much time and effort in collecting and analysing data fears that once the data are shared, others will exploit the work for their own advantage. In one case, a psychiatrist with a model predicting later schizophrenia from development before birth would not release his underlying numbers even to the writer preparing his manuscript. He sent the figures and the background only days before submission and never sent the original data, for fear that other academics would steal the idea. Many researchers also dislike being criticised. An error can be found anywhere from design to analysis to interpretation. At worst the data have no solid basis, at best the error is a human mistake, and even then it is hard to establish where what the researcher intended and what happened parted. Even without error, the same data can be read differently by different researchers, or judged unfit to answer the question they were used for, and some may avoid sharing to escape that criticism.
Economic. Economic barriers can be very important, because an over-reactive stock market can react strongly to new data. The SARS outbreak at the start of the millennium had a huge impact on the world economy. Shortly before April 2020 the Italian stock market fell 15% in a day after a statement by the head of a European economic institution, and rose 12% the next day when she said that Italy would be helped. In an environment that economically driven, news about the management of an outbreak can have a huge economic impact. The human and technical resources to prepare, annotate and communicate data, and the IT they depend on, also cost money, and under the economic pressure most companies face every day their lack is a very important limit to data transparency.
Political. Political hurdles can be very important, and the same data can be interpreted in opposite ways by people who see things from a different political point of view. Some political or ethical obstacles belong to local habits, culture and the political situation. One practitioner's example is a flu syndrome in the Arab region in 2007, which practically nobody knew about until it was used in 2020 as a model for what the coronavirus might do. In that account the countries involved did not share in order to avoid economic problems, and because they lacked trusting relationships with other countries and were not trusted by the most advanced world, which regarded them as third or fourth world. Some companies and agencies have policies that restrict data disclosure, because they paid a great deal for the data and do not want to share them. A lack of guidelines can also be a reason for data to stay locked. By 2020 clear guidelines existed at least for clinical trials, and they give instruments for moving towards transparency.
Legal. Only appropriate data that cannot identify anyone may be released, and patient privacy has to be maintained always. Where personal identifiers cannot be removed and the data fully anonymised, that is, in one practitioner's view, the only good reason not to share. Many tools for de-identifying personal data exist but may be unavailable or too expensive, which links the legal barrier to the others.
Ethical. The risks and benefits of sharing data always have to be taken into account. The start of the COVID-19 pandemic is one case. In one account, data collected in China were not all shared, and in April 2020 it was not known whether the virus had arrived in Germany or other European countries at the same time as in northern Italy, an inconsistency the account attributes to the differing ethical approaches of governments. Balancing the risks and benefits of sharing individual patient data should be the guiding ethical principle in discussions with health authorities and other agencies, which may or may not agree, while journals have their own rules.
Building a sharing capability
Where no rule requires sharing, it is up to the data holder, but a plan for individual patient data was becoming an increasingly important part of disclosure planning. Establishing data sharing practices in the organisation is important. Such practices can also enhance the image of the company or university that has them, and help an academic career.
Disclosure has consequences for an industry sponsor. For industry trials, which are usually performed to a higher standard, disclosure could be a good way to show that the sponsor is moving in the right direction and meeting patients' needs. It also generates more work, and the habit of controlling the message in interactions with colleagues in sales and marketing has to change. The scientific community, meanwhile, has to find ways of telling useful research from less useful research.
Inside a business, the obstacles that apply should be identified, experts in data transparency consulted, and a strategy for disclosing individual patient data developed and kept aligned with the regulatory documents. Medical writing and scientific communications have to act as one, and the work needs adequate resources at the business planning stage, professionals able to handle these different topics, and enough budget to make it work.
Once a researcher or medical writer has the data, they have to understand their advantages and drawbacks, the results and the possible biases, and have a clear idea of what the data say, but publishing them is no longer optional: there is no science without publishing the results.
Kinds of data that are hard to share
Several kinds of data were not yet being shared optimally in 2020. Participant data from trials were increasingly supplemented by data from patient registries, aggregation projects such as the UK Biobank and the US National Institutes of Health's All of Us programme, electronic health records and wearable devices. Those sources may improve the power of sharing, but they have limits of quality, uniformity and standardisation. Combining them with trial data may increase the risk of re-identification, because linking larger and more diverse data sets makes it easier to identify patients across them. Sharing rare-disease data is an opportunity for innovation if the challenges of patient privacy can be overcome. Aggregated rare-disease data can help make up for the paucity of patients taking part in research, for example by supplying historical controls, but most data providers exclude studies in rare diseases from sharing because of the elevated re-identification risk. Work was under way on advanced anonymisation and sharing technology, such as encryption, synthetic data modelled on the real data, and distributed analysis that takes the analytics to the data instead of sharing the data; on consent that sets out the re-identification risk and may offer an opt-out; and on alternative legal bases. Genomic data are unique to the individual and very sensitive, and their misuse could have implications for family members and future generations. Sharing them or tying them to trial data had therefore been limited to early examples such as the Psychiatric Genomics Consortium, although mechanisms to enable it were being developed. Biospecimen sharing is held back by consent, import and export rules, privacy law, re-identification risk, logistics and unclear ownership of data newly generated from a sample. Using genomic and biospecimen data responsibly, with adequately informed and consented patients and new technologies that protect their privacy, will be important to unlocking these sources.
Sharing in public health emergencies
For the most part, responsible sharing takes a deliberate approach that protects participants' privacy and consent and sponsors' confidential information and data rights, so it can be time-consuming, and some data may not be shareable at all because of confidentiality, privacy and other legal limits. That does not suit the rapid dissemination a public health emergency needs, where lack of timely access to reliable data severely hampers tracking the disease and coordinating the response. Many, the World Health Organization among them, called for sharing to be accelerated and extended during public health emergencies, a need the Ebola outbreaks in West Africa and the COVID-19 pandemic exemplified. Deficiencies in sharing basic science such as viral gene sequences, real-world data such as infection rates and symptoms, and early trial data responsibly, promptly, accurately and interoperably were identified in the Ebola outbreaks in West Africa, again with Zika, and again with COVID-19. One source dates the Ebola outbreaks in West Africa to 2014 to 2016. Another dates the epidemic to 2013 to 2016. With every outbreak ad hoc mechanisms had to be set up rather than established standards used. The World Health Organization's "COVID-19 Open" protocol for rapid online publication of research was one such mechanism, similar to approaches taken after Zika. Infrastructure and data standards for emergencies need to be set up in advance, able to overcome geopolitical, language and other barriers. In June 2020 some companies, non-profits, academia and health authorities had united across platforms to collaborate and share data more promptly, and whether that would change the paradigm after the pandemic was yet to be seen.
Solid data are the best basis for public health action during a health emergency. Effective governmental responses are possible only with accurate, real-time data. Large pools of existing data from earlier Ebola studies, the disease having been discovered in 1976, had never been fully disseminated, and much research had never been published, so when the severe outbreak triggered a rush to a cure, incomplete data led researchers down erroneous paths. During the outbreak data were collected haphazardly and not standardised, affected countries failed to communicate and were unwilling to share, and some researchers held data back until they were ready for journal publication. The post-mortem of the response identified failure to share data in a timely way as one of its major hindrances, and lack of sharing and breakdowns in communication delayed acknowledgement of the outbreak's severity and a coordinated response. When the coronavirus outbreak began, the World Health Organization declared it a public health emergency of international concern on 30 January 2020. Its Emergency Committee's statement asked the global community, in compliance with Article 44 of the International Health Regulations, to continue to show solidarity and cooperation in identifying the source of the virus and its potential for human-to-human transmission, in preparing for imported cases, and in research towards treatment. The statement noted that similar coronaviruses had required substantial efforts to enable regular information sharing and research.
In an epidemic, early data are key, used to identify the causative agents, investigate and predict spread, define diagnostic criteria and evaluate treatments and containment. Early data are valuable only if they are standardised: annotated with internationally recognised terminology such as SNOMED or LOINC, or at least captured in an agreed data model, so that data from several projects can be pooled and analysed together. The World Health Organization issued technical guidance for early investigations and a Global 2019-nCoV Clinical Characterization Case Record Form, to give member states a standardised way to collect clinical data on the disease's natural history, clinical phenotypes and treatment. With a contagion this fast, electronic case report forms were, in one 2020 account, the best way to speed the aggregation of clinical data from around the world. The maker of the data capture platform Castor offered free use of it, with electronic forms built to the WHO standard, for non-profit COVID-19 research. It reported in 2020 300 studies created, more than 200 live and over 5 million data points captured, and support for the largest global randomised trial of treatments. At the World Health Organization's consultation of September 2015 on sharing data and results during public health emergencies, international stakeholders stated that there must be timely and transparent pre-publication sharing of data and results. The organisation's recommendations that followed urged a shift from information sharing limited by publication embargoes to open sharing on fit-for-purpose pre-publication platforms, and prescribed open sharing as the default response to an emergency. They strongly encouraged sharing epidemiological and population-based data and warned of the risks of withholding data and results. The meeting also brought forward the unacceptability of non-disclosure of clinical trial data in an emergency. During COVID-19 the organisation launched a Global 2019-nCoV Clinical Data Platform for anonymised clinical data from member states. Sharing also went well beyond that platform: the release of full viral genome sequences in open databases led to rapid and reliable diagnostic tests within weeks. Outside emergencies, 12 months from study completion to public disclosure is generally considered acceptable, and the organisation recommended that this be greatly shortened in an emergency.
Re-use inside the organisation that generated the data
The value of re-using trial data was increasingly recognised in the decade to 2021, and companies sought to use the data they had already collected. Trial data have value beyond the protocol's objectives: they can inform understanding of patient populations and disease, validate new targets, and develop methods, endpoints, biomarkers and tools towards more personalised health care. Huge amounts have accumulated across industry and academia over many years. In 2020 the pandemic brought a huge interest in collaborating and sharing data across companies and academia, raising the question of what drug development could learn from it to make development cycles shorter and cheaper. External re-use was triggered by public and regulatory pressure for transparency, while internal re-use and the considerations it needs were still relatively immature in the industry in 2021.
Four areas need considering before any re-use: ethics, law and data privacy, good data science, and business sensitivity. Personal data should be used in line with participants' original expectations, as set, for example, by the secondary-use language of the informed consent form, and using anonymised or synthetic data where possible may be the best practice. Good data science means reproducible research, with analysis plans, data and results that follow the FAIR principles. For business sensitivity, data should be re-used only from locked studies or studies that have reached their primary endpoint, the project team should be made aware or involved as the programme's stage requires, and commercially confidential information has to be accounted for. Where the organisation re-using the data is the one that generated them, some of these considerations may differ slightly from external re-use.
More than 100 countries have their own data privacy laws, which may diverge, so a clear assessment of the privacy guardrails for a late-stage trial run across the globe is very challenging. A pragmatic starting point is the EU General Data Protection Regulation, still workable after Brexit because in 2021 the EU recognised the adequacy of UK privacy law, at least in the short term.
Anonymisation has no single definition, or even terminology, across the jurisdictions of a global trial. Numerous pragmatic frameworks exist: from EFSPI and PSI, PHUSE and TransCelerate, the guidance under EMA Policy 0070 and Health Canada's public release of clinical information, the UK Anonymisation Network's decision-making framework, and Khaled El Emam's publications. In 2021 the PSI and EFSPI data transparency group aimed to publish that year on anonymising clinical data for secondary use, and TransCelerate and PHUSE were working on common definitions. The EFPIA code of conduct on scientific research under the regulation was hoped to supply a harmonised industry definition when published later in 2021. A third option beside personal and anonymised data is synthetic data, generated from actual personal data but relating to no individual while keeping the statistical properties of the original, which makes them a valuable low-privacy-risk alternative for some secondary uses.
Many large pharmaceutical companies run multi-year internal re-use programmes, such as Novartis's data42, Roche's Enhanced Data and Insights Sharing and AstraZeneca's internal Data Access Policy, to give early and frictionless access to research and development data across the organisation and to make the data FAIR. Examples of re-use objectives are new insights into the drivers of response to cancer immunotherapy; a better understanding of the properties of the assessment scales used in autism studies, which changed the approach for new studies; clinical and biomarker questions answered across a broad breast cancer population, including the relationship between certain gene expression and prognosis, informing better designed studies; and identifying groups of super-responders and non-responders. Such programmes should rest on a comprehensive data governance strategy that maximises scientific insight while minimising risks such as those to privacy. Responsible re-use could then advance science in the interest of patient care, use data maximally for the public good, bring benefits to society and increase trust in organisations.
The considerations for an internal re-use strategy are grouped by the data used. With original personal data: the purpose of the re-use, definitions of primary and secondary use, wider regulatory activities, compatible use and scientific research, the informed consent form and its language on re-use, the legal basis for processing under the regulation, the ability to respond to a data subject's access request, handling consent withdrawals, matching the planned re-use against the protocol's objectives and the consent language, and decision-making. With anonymised or synthetic data: how to define anonymisation internally, an acceptable risk threshold, automation by applying standard rulesets, balancing privacy against utility, the technical and organisational measures, complexity across data modalities and linkability, the limits on what anonymised data may be used for, and the processes and assumptions for generating synthetic data. Around both sit the other considerations of a data governance strategy: a secondary-use privacy strategy, making data FAIR, good data science practice, data flow and privacy by design across the study lifecycle, systems, processes and infrastructure, internal access rules, global studies under differing privacy laws, reporting and regulatory constraints such as business sensitivity, use of data before database lock, safety signal reporting and restrictions from co-developments or divestments, other laws such as cybersecurity rules, use of samples and China's human genetic resources rules, data citizenship and codes of conduct, staff awareness and training, and adapting to changing external guidance.
Statisticians are strongly positioned to take a lead, both in spotting where secondary use would help a development programme and in collecting data so that they can be shared later. They understand the context, design and risks of a trial or molecule and the nuances of the data, knowledge that informs when data can be made available and that can be passed on as metadata to the downstream user. External data sharing involving the pharmaceutical industry grew hugely in the six to seven years to 2021 through platforms such as ClinicalStudyDataRequest.com, Vivli and TransCelerate, the pandemic brought dedicated platforms such as the COVID-19 Data Platform, and many companies were turning to their own data as an asset. Such programmes are complex and will develop over many years, and companies need to start thinking about them sooner rather than later.