Health data research in an AI-enabled world

The future belongs to researchers who care more about the problem than about their institutional silos.

AI lets researchers do far more with NHS data, and some see it as an excuse for doing whatever they choose without regard to patients. That should concern anyone whose records are in the system. A flagship project has claimed success for a model trained on patient records, yet its own preprint confirms it has no evidence of success. Patients’ choices are being worked around rather than honoured. And the body at the centre of it, HDR UK, faces no consequences while the new Health Data Research Service (HDRS, often confused with but very different to HDRUK) has funding for only five years. The rules are what make the research trustworthy, and failing legacy institutions do not wish to keep up.

Good researchers want to do everything they can to cure their chosen disease. Trustworthy Research Environments encourage them to go to the edge of what’s possible, and to expand the boundary without breaking the promises made to patients. The environment, and those who lead the research, must be robust enough to keep everyone honest and keep enthusiasm in check.

The secure environment and the project’s leadership, with data controllers, will be responsible for checking that researchers have not crossed the line. Researchers can go right up to it, and can then publish all their results to the world. Patients should see every project in the NHS App managed by NHS England, without cheating or word games that undermine trust. Use the data of those who want it used; don’t use the data of anyone who doesn’t. 

In the same way that bad money drives out good, cheating researchers drive out honest ones, and poor research drives out good. Every flow of data into, across and out of the NHS and care system should be consensual, safe, and transparent. Because there need be no conflict between good research, good ethics, and good medical care. If what an organisation says is not reliable, any consensual, safe and transparent assurances are irrelevant because they may not be true.

Cure Parkinson’s and BARBARA (the cure dementia challenge) are setting out to do what one of them says on the tin, rather than following the fickle winds of funding interests. More projects like these should see a pathway to success, and there should be a collective name for them (as the Windsors means something else, the Mitchells?). They are necessary because existing institutions get distracted by their own internal priorities and institutional politicking; the existence of the Mitchells demonstrates the fundamental failure of legacy silos.

Should the rules be abandoned for AI projects?

It’s a choice advocated by those who don’t want to follow the rules for other projects. AI allows committed researchers to do more, and is also merely the latest convenient excuse for why rules of governance and best practice shouldn’t apply to whatever a legacy silo wants to do to stay relevant.

The 2026 “Health Data Research UK” preprint has no outputs in it, because the team had not requested any outputs be checked to leave the environment before their project was paused for breaking the rules, and the project has not been submitted for approval. Since everyone agrees there were no outputs from the NHS England managed Secure Data Environment, what was the basis for the claims the research team made about success, efficacy and potential in their press release and media coverage? What was the basis for the glowing quotes from two Secretaries of State? The authors have three options: produce the checked outputs; admit the press release was based on nothing; or admit they ignored the Five Safes. Which is it? If the nominally leading institution will not respond to questions that should be trivially answerable with contemporary documentary evidence, then the other named data controllers for the project should be contacted, and all projects of those institutions given the same scrutiny, including the normal pauses on data flows for failing to comply with contractual obligations to the NHS.

When the authors claim the data is “de-identified”, that is the research world repeating the assurances of DH/E. On the other hand, the preprint has a graphic in it describing what the data is, and the “de-identified” data involved remains personal data, as medConfidential, DH/E, the paper authors, HDRUK and the ICO all agree it is. (This is why the investigation report seems not to be public).

The reliable approach is to do the full research process

AI requires a lot of computing power – the solution to questions around AI is currently to simply use even more computing power. Laziness encourages early giving up.

Medicine advances when you find something new to do and can explain what happens when you do it.. Not necessarily all the intricacies of how, but what (positive and negative). With the RECOVERY trial, something was visibly shown to work, and doctors prescribed differently to their patients. The proposal in the preprint is not for that. That is not what the preprint proposes, as the vision is to recklessly replace the doctor’s judgement, not inform it.

All the abstract testing that needs to be done for clinical safety can be done in the TRE, and once clinical safety has been measured, what leaves the environment is the new knowledge, not the large model itself. When a researcher doesn’t understand what’s going on, then a Principal Component Analysis may take some additional time and processing, it’s but less than training the AI and it’ll help the researcher write their paper in a way that a multi-gigabyte blob which may include patient data does not.

Any model that will be useful in the NHS can be tested in the TRE – it is as viable to test as it is to build the model there, allowing the extraction of safe outputs of the form “for these types of patients, dexamethasone provides better outcomes, and disproportionately so for these specific groups” as the RECOVERY preprint showed. Such an output is something that can be assessed, it can be compared, it is legible, tractable and if necessary challengeable. In short hand, the researchers have to explain themselves. HDRUK’s culture encourages none of those things when building very large models that embed patient data in the model itself in a way no one can check.

Respect patients’ data and choices

If a patient doesn’t want their data used, then that wish should be respected, not circumvented. The NHS App should show patients how data about them is used.

Addenbrooke’s admits it has a problem with snooping by its own staff. But what about the researchers? Research leaders say they fund the most curious, smart and driven people in the country, yet ask us to believe that every one of them is less curious than a random clerk at Addenbrooke’s, bored on a lunch break. That claim is simply not credible. How satisfied are the research leaders that precisely zero of their researchers went snooping on the dead girls from Southport or the boy bitten by a crocodile, or an ex-partner whose identifying characteristics are known and findable? Are research leaders as complacent as DH/E were a year ago? And if someone was caught snooping, admitted it and accepted a formal sanction under their NHS employment, what would be the consequences for their research career? The answer seems to be… none at all.

That’s not tenable, whatever the leaders of legacy research institutions wish to believe. They were the future once.

New opportunities of the Health Data Research Service

In the world we are entering, with what is now possible, some legacy institutions will insist on continuing to focus on a problem they were conceived to solve a decade ago, a problem which no longer exists. They are making themselves irrelevant, even if they remain an irritant. The future can be built on informed patient choice and empowered research, covering access for direct care and, more simply, research through the Health Data Research Service (HDRS). HDRS can be a bedrock that the Mitchell’s can build upon in a way nothing else is.

Lord Mair, Chair of the Lords Sci/Tech committee has noted with concern that HDRS is subject to a five year cliff edge; it is equally strange that HDRUK is subject to no consequence at all and more cash falls from the sky. If MRC is handing out ten year funding pots, it is perverse that HDRS doesn’t get that runway for delivery while it goes to specialists in lobbying who have failed to deliver.

Annex: HDRUK’s needs a reset – it is running out of money

medConfidential’s assessment is that HDRUK remains an active impediment to the UK’s research priorities in a changing AI-enabled world. HDRUK is culturally and institutionally set up to be one fad behind, ever chasing the old new thing. It optimises for the people and structures that feed its leadership’s egos, not for the benefit of all researchers no matter their affiliation. Research services should be equal irrespective of whether one is paying their tithe to HDRUK’s home institutions.

HDRUK mixes a leadership approach that epitomises the baboon paradox with the ethical standards of the Boris Johnson administration. Look at the wreckage of the HDRUK hubs. Look at the projects. Nothing meaningful survives any substantive scrutiny. ‘We published lots of papers’ is not the test. Which of them could only have come from HDRUK? Or complete this sentence: “Oh, HDRUK, they’re the only people who could have done ______________”… what?

The rapid evolution in AI means many of the information discovery functions which HDRUK insisted they would monopolise in 2017 would look radically different if designed today. A design for today would have no backdoors via the HDRGateway. Supposed scrutiny is a fig leaf: about as much as a teenage boy before swiping right on Tinder.

Progress will come from innovation, not those who have climbed the greasy pole in a pre-AI era. As the Royal Society Faraday Fund showed, “let’s fund young people” for speculative innovation never means to give a cheque to a young person but give a cheque to an old person, who will then tell the young person what to do. It keeps happening.

To understand the HDRUK problem, follow the money: where did the funds end up, and are they delivering what was intended? UKRI/MRC delegates assessments to HDRUK and MRC does what HDRUK suggests. There’s no one else to say “no, thanks” to the HDRUK leadership.

HDRUK is running out of money, and as always with information intermediaries, the solution to a cash crunch is to sell anything to whoever offers just enough to postpone the immediate crisis.

medConfidential welcomes the public agreement from the President of the Academy of Medical Sciences with its assessment of the failure of parts of the regional SDE programme, parts that HDRUK was supposed to deliver and a mess that HDRS has to tidy up safely.

The President of AcMedSci (whose main other job is being Director of HDRUK) has decided that HDRUK’s response should be to abandon all sovereignty and send the NHS data to the US for use in a “safe environment” outside of UK law, principally because HDRUK is running out of cash (US organisations giving HDRUK cash is a tax write-off in the US). Such an abandonment is incompatible with efforts to increase UK sovereignty, and is incoherent in the face of the pathway to success for the Health Data Research Service. 

Given the cash and unsupervised data that HDRUK flows to leadership’s home institutions in Scotland, has Research Data Scotland (being “advised” by HDRUK) chosen to send Scottish patient data to the Mayo in return for cash? Or does the baboon show a different end on either side of the border?

HDRUK is not the first research charity to implode this way, as Turing has before it, akin to how UK Biobank’s catastrophe happened. The clear (possibly deliberate) confusion between HDRS and HDRUK means HDRUK must change its name and undergo a full reset; any sockpuppets that survive will be those who demonstrate value in alternate homes. 

==

Find out what happens next: Sign up for our newsletter (we don’t email often) or get small frequent updates via Substack — free to follow, and we are grateful to all those who can donate to help more of this work.

Leave a Reply

Your email address will not be published. Required fields are marked *