McKay et al.

BMC Medical Ethics (2023) 24:49 BMC Medical Ethics

https://doi.org/10.1186/s12910-023-00927-8

DEBATE Open Access

Artificial intelligence and medical research

databases: ethical review by data access
committees
Francis McKay1* , Bethany J. Williams2, Graham Prestwich3, Daljeet Bansal2, Darren Treanor2,4,5,6 and
Nina Hallowell7

Abstract
Background It has been argued that ethics review committees—e.g., Research Ethics Committees, Institutional
Review Boards, etc.— have weaknesses in reviewing big data and artificial intelligence research. For instance, they
may, due to the novelty of the area, lack the relevant expertise for judging collective risks and benefits of such
research, or they may exempt it from review in instances involving de-identified data.
Main body Focusing on the example of medical research databases we highlight here ethical issues around de-iden-
tified data sharing which motivate the need for review where oversight by ethics committees is weak. Though some
argue for ethics committee reform to overcome these weaknesses, it is unclear whether or when that will happen.
Hence, we argue that ethical review can be done by data access committees, since they have de facto purview of big
data and artificial intelligence projects, relevant technical expertise and governance knowledge, and already take
on some functions of ethical review. That said, like ethics committees, they may have functional weaknesses in their
review capabilities. To strengthen that function, data access committees must think clearly about the kinds of ethical
expertise, both professional and lay, that they draw upon to support their work.
Conclusion Data access committees can undertake ethical review of medical research databases provided they
enhance that review function through professional and lay ethical expertise.
Keywords Artificial intelligence, Medical research databases, Health data repositories, Data access committees,
Research ethics committees, Ethical review, Ethical expertise, Public involvement

7
*Correspondence: The Ethox Centre and the Wellcome Centre for Ethics and Humanities,
Francis McKay Nuffield Department of Population Health, University of Oxford,
francis.mckay@newcastle.ac.uk Oxford OX3 7LF, UK
1
Population Health Sciences Institute, University of Newcastle, NE2
4AX Newcastle Upon Tyne, UK
2
National Pathology Imaging Co‑operative, Leeds Teaching Hospitals
NHS Trust, Leeds LS9 7TF, UK
3
Yorkshire and Humber Academic Health Science Network, Unit 1, Calder
Close, Calder Park, Wakefield WF4 3BA, UK
4
Department of Pathology, University of Leeds, Leeds, UK
5
Department of Clinical Pathology, Linköping University, Linköping,
Sweden
6
Center for Medical Image Science and Visualization (CMIV), Linköping
University, Linköping, Sweden

© The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. The Creative Commons Public Domain Dedication waiver (http://​creat​iveco​
mmons.​org/​publi​cdoma​in/​zero/1.​0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
McKay et al. BMC Medical Ethics (2023) 24:49 Page 2 of 7

Background expertise they solicit to guide the review process. Most

Medical research databases—collections of health infor- notably, DACs should be mindful to include independent
mation stored for the purpose of research—are an impor- ethical experts, both professional (e.g., bioethicists, data
tant mechanism by which artificial intelligence (AI) is ethicists, etc.) and lay (in the form of patient and public
trained on healthcare data [1–3].1 Databases may contain involvement [PPI]).
identifiable patient information and, depending on the To make our case we examine the ethical challenges of
type in question, could hold a variety of medical informa- sharing de-identified medical data via research databases.
tion, for instance, genomic data, electronic care records, In doing so, we distinguish between two types of chal-
medical images, etc. To protect individual data subjects’ lenges: “upstream” ethical issues impacting individual
confidentiality, however, data controllers generally de- data subjects’ interests (and which are well protected by
identify data prior to sharing it for research. Insofar as existing governance mechanisms), and “downstream”
de-identifying data is thought to reduce risks to patient issues impacting collective interests (and which are less
privacy, it also limits the need, in many jurisdictions, for well protected). Following that, we make the case for
ethical review prior to distribution [8–11]. DACs as a strategic mechanism for ethical review regard-
Sharing de-identified medical data for AI research ing the latter, highlighting the importance of independent
can still raise ethical concerns, however. Studies have ethical expertise in the process.
shown [12–15], for instance, that multiple ethical issues
can arise as a result of the downstream applications of Main text
AI. Consequently, there is need for ethical oversight of Ethical challenges of sharing de‑identified data
research using data from medical research databases, Multiple jurisdictions (e.g., the United Kingdom, the
even when data is de-identified.2 United States, Australia, the Netherlands) permit de-
Seemingly, ethics review committees—e.g., Research identified data sharing from medical research databases
Ethics Committees (RECS) in the UK, Institutional with limited or exempted ethical review [9, 13, 17].
Review Boards (IRBs) in the USA, etc.—are well placed As Scott et al. note, such review can include multiple
to undertake this ethical oversight. As Ferretti et al. have approaches, such as “informing the committee of the pro-
argued, however, ethics committees may be “function- ject (e.g. by submitting an exemption form) but not sub-
ally weak” in reviewing big data and AI research due to mitting an application for review”; “some form of partial
commonly lacking expertise for assessing the collective or expedited review”; or “bypassing review by an ethics
risks and benefits of such research [10]. In addition, they committee completely before commencing the research
may have “purview weaknesses” as exemptions to ethical project” [9]. The reasons for this are well grounded from
oversight are often available for projects using de-identi- a research ethics perspective. When data is de-identified,
fied data. Several authors have therefore argued for ethics it is understood to reduce risks to data subjects’ privacy
committee reform, so that they might develop purview [8]. Insofar as a core purpose of research ethics oversight
and functional strengths for reviewing such research is to protect such interests [18, 19], the need for review is
[10, 11, 16]. Pending any reforms, we ask: how else might often seen to be obviated.
ethical oversight be provided? We argue here that such Many have questioned the degree to which de-iden-
work can be done by data access committees (DACs). tification protects data subjects in an era of big data
DACs, we argue, are an appropriate mechanism for ethi- due to the potential for re-identification [3, 8, 20]. Even
cal review of research applications to medical research where data is robustly de-identified, however, research-
databases as they have de facto purview of big data and ers intending to use such data may still encounter ethi-
AI projects alongside relevant technical expertise and cal issues in their research. For instance, secondary uses
governance knowledge. As will be shown, they also often of data often entail ethical issues around collective rather
take on functions of ethical review. However, like eth- than individual interests [6, 16, 18]. Hence, it has been
ics committees, they may have functional weaknesses in argued that de-identified medical data sharing should
relation to ethical review. To strengthen that function, we be governed by a public health framework, in which the
suggest data access committees must think clearly about focus is on maximising public benefit and minimising
their membership structures, and the kinds of ethical collective harms, as opposed to a research ethics frame-
work in which the focus is on individual risk protection
through consent forms, confidentiality agreements, per-
1
“Research database” is the term favoured by the UK’s Health Research sonal information sheets, etc., and which are the spe-
Authority [4]. However, they may go under a variety of other names, such as cialty of RECs [18].
health data repositories, registries, databanks, datalakes, etc. [3, 5–7]. The downstream risks of de-identified medical data
2
We focus here on ex ante or anticipatory ethical review, though much of
what can be said would also apply to ex post review.
sharing depend on a variety of factors, such as the kind of
McKay et al. BMC Medical Ethics (2023) 24:49 Page 3 of 7

data (including linked data) being shared, who it is shared Ethical oversight through DACs
with, the purposes for which it is shared, and the specific To maximize benefits of data sharing while protecting
community interests affected by the research outcomes. against downstream harms, ethical review of research
That said, general issues have been raised in recent years, applications to medical research databases is needed.
particularly around the possibilities of commercialisation By “ethical review” we mean oversight through ethical
and bias. reflection by persons with some form of “ethical exper-
In relation to the sharing of pathology image data, for tise” but no conflict of interest in the research.
instance, such research could exacerbate health inequali- Ordinarily, ethics review committees provide that over-
ties through the privatisation of diagnostic knowledge sight. As mentioned, however, de-identified data sharing
and technologies [21]. This is a possibility for medical from medical research databases may be exempt from
AI research in general insofar as researchers are com- ethical review due to the limited risks to individual inter-
monly expected to be commercial organisations, includ- ests that anonymity brings.3 Even when ethics commit-
ing large technology companies, since they are the ones tees undertake review of research projects, however, they
who possess the advanced engineering expertise and may still have what Ferretti et al. call “functional weak-
technical resources for bringing medical AI research nesses” in reviewing big data and AI research [10]. Ethics
to market and to clinic in a usable form. As Mittelstadt committees are strong when undertaking ex ante review
points out, however, there can be tensions between medi- of traditional biomedical research (e.g., clinical trials,
cal and commercial AI research due to the lack of “com- longitudinal cohort studies, etc.). Here research subjects
mon aims and fiduciary duties” [21]. Whereas medicine are directly involved in data collection, demanding ethi-
has long developed professional duties and codes of con- cal protections in the form of consent, confidentiality,
duct to facilitate the goal of improving health, the same is etc. Research ethics committees have a strong history
not necessarily true of AI, especially in the private sector, of protecting human subjects’ interests in this way. They
where profit incentives may provide a conflict of interest are less well equipped, however, for overseeing potential
with public health benefits and where AI may, as Spector- downstream harms, largely due to a perceived lack of rel-
Bagdady shows, not be held to the same standard of regu- evant expertise for judging collective benefits and risks
lation [6]. for big data and AI based research [10, 16].
The potential for monopolies and unequal distribu- The above weaknesses suggest an “ethics oversight gap”
tion of healthcare as a result of proprietary knowledge for big data and medical AI research. Responding to that
and tools is not the only avenue for downstream health gap, several authors have argued for the need for ethics
inequalities, for an additional way is through bias. Often committee reform [10]. Since it is not clear whether and
AI bias is understood as datasets that are unrepresenta- when such reforms may happen, however, we argue that
tive in terms of relevant demographic and clinical crite- data access committees (DACs) could provide an alterna-
ria (e.g., health status, age, ethnicity, gender, geographical tive site for that ethical review. Unlike ethics committees,
location, etc.). Multiple authors have noted the impact DACs generally have technical expertise and governance
bias may have for community interests [22, 23]. As Hao knowledge around data sharing, making them well-suited
notes, however, “bias can creep in long before the data for navigating the growing complexities of big data and
is collected as well as at many other stages of the deep- AI research. Moreover, as data access managers, requests
learning process” [24]. For instance, it can be found in for de-identified medical data come through them by
how researchers frame the questions, problems, or cat- default. Though not all DACs operate the same way,
egories of research, in how they determine cohort selec- empirical research by Shabani et al. also suggests that
tions in a database, in how they decide what linked data is many DACs already take ownership of a range of ethi-
relevant, and in how they adjust an algorithm’s weights to cal duties, including providing oversight of downstream
improve its predictive accuracy. For instance, Obermeyer ethical issues by restricting or flagging culturally or
et al. [25] found that a US healthcare algorithm, which politically controversial uses of data (such as those going
used historical healthcare costs as a proxy for healthcare counter to prevailing social norms) [27, 28]. In summary,
needs, wrongly classified African Americans as being at then, DACs are an appropriate site for review, insofar as
lower risk due to having historically lower costs when they have de facto purview, functional strengths in the
compared to white Americans. The system affected mil- governance of big data and AI research generally, and
lions of people and perpetuated racial prejudice by label-
ling African Americans as requiring less medical care.
Though the data here was to some degree biased, so too
was the framing of the research and was only improved 3
As Bernstein et al. point out [26], such exclusions may apply whether data
is de-identified or not, as in the case of IRBs in the USA, where such assess-
when the proxy measure was removed. ments are considered outside the purview of the board.
McKay et al. BMC Medical Ethics (2023) 24:49 Page 4 of 7

because they often already take on informal responsibili- [36]. Related to that, there is also what one might call a
ties for ethical review. “participation deficit” for lay contributors in medical AI
Since not all DACs operate the same way, however, research generally, partly due to the novelty of its appli-
there is need for general advocacy for medical DACs cation, which mirrors a more general participatory gap
to take on responsibilities for ethical review where that regarding the use of AI in society [37].
is not already the case. Insofar as they do that, it is also Though we recognize that public understanding and
important to note the need for strengthening the review engagement around medical AI may be constrained due
process. This is because, like RECs, DACs face limita- to the novelty of its application in society, this does not
tions in their capacity for ethical review. As Cheah and mean, however, that lay participation on DACs should be
Piasecki put it, DACs are in danger “of underestimating overlooked. The benefits of lay representatives in health
or misidentifying potential harms to data subjects and research are well known [38], suggesting a prima facie
their communities” as they “do not necessarily know duty for their inclusion on DACs. Indeed, Health Data
what these group harms will be” [29]. Given potential Research UK have argued that this should be standard
functional weaknesses in terms of judging downstream practice [39]. That said, the lack of clarity around what
risks and benefits, DACs should be mindful that they constitutes lay AI ethics expertise or the relevance of lay
seek relevant ethical expertise to guide their reflections. members to nuanced decision making around data shar-
The question of what constitutes ethical expertise is ing for AI research means further justification is needed.
a long standing one [30–32]. Though research has not Hence, we highlight here important procedural and sub-
explicitly addressed that question vis-à-vis medical AI stantive justifications in relation to medical AI research.
research, it has made the case for such expertise in medi-
cal research generally. Inspired by that work, we recog- Public involvement on DACs
nise that ethical expertise is possible and desirable for AI PPI can have value for evidencing procedural fairness
research, and that such expertise can take multiple forms, insofar as it includes healthcare stakeholders in the deci-
including independent professional ethicists (e.g., bioeth- sion-making processes around health. This procedural
icists, legal scholars, critical data scholars, social scien- value is important for the ethical oversight of research
tists, etc.) and lay stakeholders (e.g., PPI).4 Murtagh et al. databases, for if one of the goals of DACs is to maxim-
[34] have discussed the relevance such interdisciplinary ise the utility of data for public benefit, the question of
expertise has for responsible data sharing, arguing that, what constitutes public benefit is one that, procedurally,
since big data and AI projects are complex, they require requires broad public deliberation to determine. Multiple
a variety of disciplinary and non-disciplinary experts to mechanisms exist for deliberating about collective val-
fulfil important roles such as providing understandings ues for AI, from online crowd sourcing to citizen forums
of laws and regulations (in the case of legal scholars), or [40–42]. Public involvement is a long-standing comple-
highlighting relevant contextualising factors that impact mentary mechanism to those processes and can continue
research trajectories (in the case of social scientists), etc. to be useful within the specific local contexts of research
DACs need to specify for themselves the expertise they applications. Lay representation on DACs is one way
require for making judgments about data access requests. that deliberation can occur for the sake of medical data
Prima facie, however, the inclusion of lay participants sharing.
may appear more challenging than professional contribu- Procedural fairness in decision-making may also have
tors. The reason comes from the status of professional implications for the trustworthiness of database research.
versus lay expertise in relation to medical AI research As Kerasidou notes, trustworthiness means an individ-
generally. Although there is a robust body of professional ual or organisation evidencing that their decisions are
research in critical algorithm studies and medical AI in the best interests of another [43]. Trustworthiness
ethics [14, 35] legitimizing the notion that an emergent thus shifts the burden of public confidence away from
disciplinary expertise exists on the ethics of medical AI, the public toward organisations, which must prove they
there is a relative lack of public awareness and under- are worthy of trust. Trustworthiness in regard to medi-
standing of AI and of its relevance to health research cal AI research can be shown by developing algorithms
that are lawful, ethical, and robust [44]. Public partici-
pation in the reviewing process can further engender
4
Downstream impacts of research may be impacted as much by research that trustworthiness by ensuring that representatives
quality as research ethics. For instance, a study by Wong et al. [33] shows with shared public interests have voice in the decision-
how a poorly validated proprietary sepsis prediction model implemented in
the US predicted onset of sepsis in a way that was “substantially worse” than making process (which is important where commercial
was reported by its developer. Hence, ethical reflection may also be sup- involvement gives reason to question the priorities of
ported by a range of disciplinary expertise in AI and medicine, and not just the researchers involved). It thereby provides confidence
professional ethicists or lay contributors.
McKay et al. BMC Medical Ethics (2023) 24:49 Page 5 of 7

that downstream collective interests have been taken application. The example of smaller research groups is
into consideration, which relates to PPI’s substantive a special case, requiring further deliberation. That issue
contributions. notwithstanding, appropriately resourced DACs, i.e.,
PPI has substantive value for medical AI research those associated with research consortia or institutes,
insofar as it can explicate potential collective interests can nonetheless provide a profitable means for filling in
at stake in research applications. Such contributions are the gap left by ethics committees.
relative, however, to the different knowledge and experi- It may also be argued that there are alternatives ways
ence PPI members bring. At its broadest level, this means in which to address the ethics oversight gap, which could
patients and publics. According to Fredriksson and Trit- obviate the need for DACs in this regard. Bernstein
ter, patients and publics, though often conflated, bring et al. [26] provide a promising example of an alternative
distinct contributions to PPI discussions: patients offer- approach in what they call an “Ethics and Society Review
ing “sectional” insights based on their experiences as board,” which is an ad hoc panel devised at Stanford Uni-
health service users, publics providing a broader societal versity for the purpose of reviewing the downstream
perspective based on their civic understanding [45]. impacts of research. Additionally, ethics awareness
When examined more closely, however, it becomes training might be provided to raise AI developers’ ethi-
apparent that PPI members represent a diverse range cal competencies so that it can inform their work. For-
of subject positions and collective interests. These may tunately, building ethical reflection into the medical AI
include the general interests of patients as a whole, the research pipeline is not an either/or situation. Moreover,
specific interests of particular patient communities (such given the complexity of medical AI research, as well as the
as cancer patients), the interests of community groups multiple contexts in which it occurs (involving universi-
defined by demographic characteristics (such as ethnic- ties, hospitals, and commercial organisations to different
ity, age, or gender), and the broader interests of citizens degrees), it is desirable to have multiple means available.
[40]. The reason they can do this is because they have DACs, however, are complementary to that and are well
cultural knowledge about collective interests, which placed for providing formal oversight (of the kind usu-
provides them with acquired vigilance allowing them ally reserved for RECs) insofar as they are already well-
to anticipate relevant community harms and benefits. established, have purview over large amounts of medical
This vigilance regarding community risks and benefits AI research globally, have strengths in governance and
means they are well placed to anticipate how novel forms data sharing, and already take on some functions of ethi-
of research may impact them. Regarding applications cal review. They therefore could be more easily adapted
to medical research databases, such “ethical expertise” and capitalised on, when contrasted with ad hoc panels,
regarding anticipatory harms and benefits can be used to which would take time to implement at scale.
reflect on possible community impacts, clarify commu- Regarding bringing lay representation onto DACs, that
nity needs and preferences, and thus guide researchers in could also raise similar kinds of challenges that have been
how to avoid them. discussed about PPI in other areas. For instance, there is
the impracticality of representing all community inter-
Further questions
ests. There is a variety of stakeholders to AI, but it would
There are challenges, however, with DACs taking on be unrealistic to survey all interest groups for every data
the functions of ethics review and of the role of PPI in access request. DACs would have to determine mem-
that process. It may be argued, for instance, that not bership structures for themselves, though we recognise
all DACs possess sufficient capacity for such review. that there will be times when broad patient and public
Some smaller research groups, for instance, comprise involvement will suffice to provide oversight on collec-
of only one or two people, such as a PI or a post-doc- tive interests, and there will be times when more specific
toral researcher, who manage the data access requests community input will be needed. It would be the respon-
[46, 47]. Such groups may lack the ability to outsource sibility of DACs to be alert to the kind of help that is
ethical review and there could be a conflict of inter- needed and when.
est if members undertook the review themselves. Another issue concerns how to mediate conflict of
Such groups may benefit from alternative approaches, interests. It is possible PPI members will understand ben-
therefore. It has been suggested [48], for instance, that efits and risks differently. Here we suggest that diversity
central DAC infrastructures could be developed via of viewpoints does not preclude publics from reaching
funding agencies to provide that support. Alternatively, compromise. Insofar as PPI representatives inhabit mul-
ethical review for smaller DACs might be shifted on to tiple subject positions, they are able to move beyond sec-
the data applicant, who would provide evidence of hav- tional interests and recognise the need for trade-offs in
ing gone through independent ethical review prior to the service of a wider public good.
McKay et al. BMC Medical Ethics (2023) 24:49 Page 6 of 7

Perhaps most importantly is the advisory nature of for publication. All authors have approved the submitted version and are
accountable for the contributions and integrity of the paper.
public involvement in general, which often entails the
possibility for “involvement tokenism,” that is, using PPI Funding
as a check box exercise. What “good” public involvement The authors of this study are members of the National Pathology Imaging
Co-operative, NPIC (Project no. 104687) which is supported by a £50 m invest-
looks like is an ongoing research question [49, 50]. To ment from the Data to Early Diagnosis and Precision Medicine strand of the
guard against the ever-present possibility of tokenism, government’s Industrial Strategy Challenge Fund, managed and delivered by
however, we suggest DACs provide opportunities for UK Research and Innovation (UKRI). NH’s research is funded by the Li Ka Shing
Foundation. Both NH and FM are also members of the Ethox Centre and the
devolved decision-making to PPI members, for instance, Wellcome Centre for Ethics and Humanities which is supported by funding
by ensuring all members, including lay members, carry from the Wellcome Trust (Grant no 203132). No specific funding was received
equal weight when deciding whether applications pro- for this study. For the purpose of Open Access, the author has applied a CC BY
public copyright licence to any Author Accepted Manuscript (AAM) version
ceed, are rejected, or are sent back for revision. arising from this submission.

Availability of data and materials

No new data were created during this study.
Conclusion
Medical research databases are an important means by
Declarations
which AI is trained on health data.Given that researchers
may face ethical issues in the application of their work, Ethics approval and consent to participate
pre-emptive ethical oversight of research applications is Not applicable.
important. Ethics review committees, however, may lack Consent for publication
purview or functional strengths when it comes to review- Not applicable.
ing big data and AI based medical research. In lieu of eth-
Competing interests
ics committee reforms, DACs are a viable alternative, and DT is director of a research program (NPIC) which includes in-kind funding
are in some ways better placed than ethics committees from industry partners including Leica Biosystems, Roche, and Sectra. DT has
due to de facto purview strengths, technical and govern- no other conflicts to declare. All other authors have no conflict of interest to
declare.
ance expertise, and general duties for scientific and ethi-
cal review. That said, like RECs, DACs may still exhibit
potential functional weakness in their capacity for ethi- Received: 25 May 2022 Accepted: 22 June 2023
cal expertise. Hence, it is recommended that they solicit
input in the form of professional and lay ethical experts
to strengthen that function. The inclusion of lay partici-
References
pants may appear more challenging than professional
1. Health Data Research Innovation Gateway. 2020. https://​www.​healt​hdata​
contributors, however, due to a lack of public awareness gatew​ay.​org/. Accessed 9 Dec 2021.
and understanding of AI and a general “participation def- 2. Cimino JJ, Ayres EJ. The clinical research data repository of the US
National Institutes of Health. Stud Health Technol Inform. 2010;160(Pt
icit” for lay contributors in medical AI research. None-
2):1299–303.
theless, lay members should continue to be an important 3. Caplan A, Batra P. The ethics of using de-identified medical data for
cornerstone of ethical reflection due their procedural and research without informed consent. Voices Bioethics. 2019;5:1–5.
4. National Research Ethics Service. Ethical review of research databases.
substantive contributions.
2010. https://​www.​hra.​nhs.​uk/​docum​ents/​376/​ethic​al-​review-​of-​resea​
rch-​datab​ases.​pdf. Accessed 13 Jul 2021.
5. Lee Y-J. Welch Medical Library Guides: Finding Datasets for Secondary
Abbreviations Analysis: Research Data Repositories & Databases. 2021. https://​browse.​
AI Artificial intelligence welch.​jhmi.​edu/​datas​ets/​nih-​data-​repos​itori​es. Accessed 18 Jan 2022.
DACs Data access committees 6. Spector-Bagdady K. Governing secondary research use of health data and
IRBs Institutional Review Boards specimens: the inequitable distribution of regulatory burden between
PPI Patient and Public Involvement federally funded and industry research. J Law Biosci. 2021;8:lsab008.
RECs Research Ethics Committees 7. PathLAKE. PathLAKE Research Database - Privacy Notice. 2020. https://​
www.​pathl​ake.​org/​wp-​conte​nt/​uploa​ds/​2020/​09/​PathL​AKE-​Resea​rch-​
Acknowledgements Datab​ase-​priva​cy-​notice-​v1.0-​090920.​pdf. Accessed 5 Aug 2021.
Thanks go to the members of the National Pathology Imaging Cooperative, 8. Rothstein MA. Is deidentification sufficient to protect health privacy in
its patient and public advisory group, the University of Oxford’s Ethox Centre, research? Am J Bioeth. 2010;10:3–11.
and the Yorkshire and Humber Academic Health Science Network for ongoing 9. Scott AM, Kolstoe S, Ploem MCC, Hammatt Z, Glasziou P. Exempting
discussions and support on the ethics of digital pathology and medical AI low-risk health and medical research from ethics reviews: comparing
research. Australia, the United Kingdom, the United States and the Netherlands.
Health Res Policy Sys. 2020;18:11.
Authors’ contributions 10. Ferretti A, Ienca M, Sheehan M, Blasimme A, Dove ES, Farsides B, et al.
FM conceived of the work; acquired, analysed and interpreted the data; Ethics review of big data research: what should stay and what should be
drafted the initial submission; and revised it for publication. BW, GP, DB, DT, reformed? BMC Med Ethics. 2021;22:51.
NH helped to interpret the data, draft the initial submission, and revise it
McKay et al. BMC Medical Ethics (2023) 24:49 Page 7 of 7

11. Friesen P, Douglas-Jones R, Marks M, Pierce R, Fletcher K, Mishra A, et al. 38. Staniszewska S. Patient and public involvement in health services and
Governing AI-Driven Health Research: Are IRBs Up to the Task? Ethics health research: a brief overview of evidence, policy and activity. J Res
Hum Res. 2021;43:35–42. Nurs. 2009;14:295–8.
12. Fullerton SM, Lee SS-J. Secondary uses and the governance of de- 39. Health Data Research UK. Building trust in data access through public
identified data: lessons from the human genome diversity panel. BMC involvement in governance: survey findings and recommendations from
Med Ethics. 2011;12:16. HDR UK’s Public Advisory Board. UK: Health Data Research; 2021.
13. Ienca M, Ferretti A, Hurst S, Puhan M, Lovis C, Vayena E. Considerations 40. McKay F, Williams BJ, Prestwich G, Treanor D, Hallowell N. Public govern-
for ethics review of big data health research: a scoping review. PLoS ONE. ance of medical artificial intelligence research in the UK: an integrated
2018;13:e0204937. multi-scale model. Res Involv Engagem. 2022;8:21.
14. Morley J, Machado CCV, Burr C, Cowls J, Joshi I, Taddeo M, et al. The ethics 41. O’Doherty K, Einsiedel E. Public engagement and emerging technologies.
of AI in health care: a mapping review. Soc Sci Med. 2020;260:113172. Vancouver: UBC Press; 2013.
15. Vayena E, Blasimme A, Cohen IG. Machine learning in medicine: address- 42. Rahwan I. Society-in-the-loop: programming the algorithmic social
ing ethical challenges. PLoS Med. 2018;15:e1002689. contract. Ethics Inf Technol. 2018;20:5–14.
16. Samuel G, Chubb J, Derrick G. Boundaries between research ethics and 43. Kerasidou A. Trust me, I’m a researcher!: the role of trust in biomedical
ethical research use in artificial intelligence health research. J Empir Res research. Med Health Care and Philos. 2017;20:43–50.
Hum Res Ethics. 2021;16:1–13. 44. Hleg AI. Ethics guidelines for trustworthy AI: independent high-level
17. El Emam K, Rodgers S, Malin B. Anonymising and sharing individual expert group on artificial intelligence. Brussels: European Commission;
patient data. BMJ. 2015;350:h1139. 2019.
18. Ballantyne A. Adjusting the focus: a public health ethics approach to data 45. Fredriksson M, Tritter JQ. Disentangling patient and public involvement
research. Bioethics. 2019;33:357–66. in healthcare decisions: why the difference matters. Sociol Health Illn.
19. Hemminki E. Research ethics committees in the regulation of clinical 2017;39:95–111.
research: comparison of Finland to England, Canada, and the United 46. Shabani M, Dyke SOM, Joly Y, Borry P. Controlled access under review:
States. Health Res Policy Sys. 2016;14:5. improving the governance of genomic data access. PLoS Biol.
20. Sweeney L. Simple Demographics Often Identify People Uniquely. 2000. 2015;13:e1002339.
https://​datap​rivac​ylab.​org/​proje​cts/​ident​ifiab​ility/​paper1.​pdf. Accessed 47. Shabani M, Borry P. “You want the right amount of oversight”: interviews
18 May 2022. with data access committee members and experts on genomic data
21. Mazer BL, Paulson N, Sinard JH. Protecting the pathology commons in access. Genet Med. 2016;18:892–7.
the digital era. Arch Pathol Lab Med. 2020;144:1037–40. 48. Shabani M, Knoppers BM, Borry P. From the principles of genomic data
22. Ibrahim H, Liu X, Zariffa N, Morris AD, Denniston AK. Health data poverty: sharing to the practices of data access committees. EMBO Mol Med.
an assailable barrier to equitable digital health care. Lancet Digit Health. 2015;7:507–9.
2021;3:e260–5. 49. Liabo K. Public involvement in health research: what does ‘good’ look like
23. Kaushal A, Altman R, Langlotz C. Geographic distribution of US cohorts in practice? Res Involv Engagem. 2020;6:1–12.
used to train deep learning algorithms. JAMA. 2020;324:1212. 50. McCoy MS, Jongsma KR, Friesen P, Dunn M, Neuhaus CP, Rand L, et al.
24. Hao K. This is how AI bias really happens—and why it’s so hard to fix. National standards for public involvement in research: missing the forest
2019. https://​www.​techn​ology​review.​com/​2019/​02/​04/​137602/​this-​is-​ for the trees. J Med Ethics. 2018;44:801–4.
how-​ai-​bias-​really-​happe​nsand-​why-​its-​so-​hard-​to-​fix/. Accessed 18 May
2022.
25. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias Publisher’s Note
in an algorithm used to manage the health of populations. Science. Springer Nature remains neutral with regard to jurisdictional claims in pub-
2019;366:447–53. lished maps and institutional affiliations.
26. Bernstein MS, Levi M, Magnus D, Rajala B, Satz D, Waeiss C. ESR: ethics and
society review of artificial intelligence research. 2021.
27. Shabani M, Dove ES, Murtagh M, Knoppers BM, Borry P. Oversight of
genomic data sharing: what roles for ethics and data access committees?
Biopreserv Biobank. 2017;15:469–74.
28. Shabani M, Thorogood A, Borry P. Who should have access to genomic
data and how should they be held accountable? Perspectives of
Data Access Committee members and experts. Eur J Hum Genet.
2016;24:1671–5.
29. Cheah PY, Piasecki J. Data access committees. BMC Med Ethics.
2020;21:12.
30. Weinstein BD. The possibility of ethical expertise. Theoret Med.
1994;15:61–75.
31. Rasmussen LM. Ethics expertise: history, contemporary perspectives, and
applications. Dordrecht: Springer; 2005.
32. Yoder SD. The nature of ethical expertise. Hastings Cent Rep.
1998;28:11–9.
Ready to submit your research ? Choose BMC and benefit from:
33. Wong A, Otles E, Donnelly JP, Krumm A, McCullough J, DeTroyer-
Cooley O, et al. External validation of a widely implemented proprietary
• fast, convenient online submission
sepsis prediction model in hospitalized patients. JAMA Intern Med.
2021;181:1065–70. • thorough peer review by experienced researchers in your field
34. Murtagh MJ, Blell MT, Butters OW, Cowley L, Dove ES, Goodman A, et al. • rapid publication on acceptance
Better governance, better access: practising responsible data sharing in
• support for research data, including large and complex data types
the METADAC governance infrastructure. Hum Genomics. 2018;12:12–24.
35. Gillespie T, Seaver N. Critical Algorithm Studies: a Reading List. Social • gold Open Access which fosters wider collaboration and increased citations
Media Collective. 2016. https://​socia​lmedi​acoll​ective.​org/​readi​ng-​lists/​ • maximum visibility for your research: over 100M website views per year
criti​cal-​algor​ithm-​studi​es/. Accessed 25 Jan 2022.
36. Castell S, Robinson L, Ashford H. Future data-driven technologies and the At BMC, research is always in progress.
implications for use of patient data. London: Ipsos Mori; 2018.
37. RSA. Artificial intelligence: real public engagement. London: RSA; 2018. Learn more biomedcentral.com/submissions