AI and the NHS: is it the silver bullet that will improve the health service’s productivity?

Please note that views expressed in guest articles on our website are the authors' own and do not necessarily reflect the views of the Nuffield Trust.

76 years since its creation, the National Health Service (NHS) remains one of the most beloved and valuable institutions in the UK. However, it is evident that in 2024, the NHS is struggling. The challenges facing the health service are so significant, with increasing expenditure and worsening outcomes, that minor policy ‘tweaks’ are unlikely to be sufficient. Instead, there is increasing pressure on the new government to try something bolder to change the NHS’s modus operandi (MO) to increase its productivity and improve its outcomes.

The previous government believed that artificial intelligence (AI) held this necessary MO-changing power. Several of their policy documents espoused the intention to harness AI’s potential to make health care more efficient, more effective and more sustainable, as well as more evidence-based, less wasteful and less harmful. The ultimate hope being that by creating these informational feedback loops, the NHS as a whole would learn how to improve the experience of care, improve population health and reduce per capita costs (the ‘triple aim'). Now, a month into power, it seems that the new Labour government is also planning to put many of its metaphorical eggs into the AI basket.

Wes Streeting, the Secretary of State for Health and Care, and Peter Kyle, the Secretary of State for Science, Innovation and Technology, have been clear about their intentions to leverage data and AI to “build an NHS fit for the future”, turn the UK into a life sciences and medical superpower, and to use the NHS’s data to boost the economy. In many ways this continued focus on the opportunities represented by NHS data and AI is welcome. AI has been shown to be theoretically capable of predicting, diagnosing and ‘treating’ disease as well as automating back-end operational tasks, and could therefore take considerable pressure off front-line services. Yet it is crucial that this technological optimism is appropriately tempered with realism before the AI hype train carries the NHS to an unintended (and undesirable) destination.

What must be taken into consideration?

First, there are practical considerations related to the limitations of the NHS’s existing legacy information infrastructure. AI development requires access to large volumes of patient data. Not only are these data often highly sensitive, but they are currently stored in hundreds of siloed databases controlled by different data controllers, in inconsistent formats, and governed by different access rules.

Furthermore, patient data were never designed to be used for AI development purposes: their primary function is to act as an aide-mémoire for clinicians when dealing with the direct care of the patient. Consequently, patient datasets are far from perfect. Instead, patient datasets can be fraught with errors and omissions. Lastly, for AI to be useful, it must be seamlessly integrated into live clinical systems used at the point of care. NHS clinical systems are currently plagued by a lack of interoperability. Rarely do the same systems record the same information in the same format, and even more rarely do the different systems ‘talk to each other’. It is not unusual, therefore, for clinicians to have to log in to multiple systems just to manage the care of one patient.

Second, there are regulatory considerations related to the need to ensure that AI used in the NHS is safe and effective. Currently, it is unclear how exactly to validate and evaluate the performance of an AI system intended to be used for clinical care. Very few AI systems have been tested in a clinical trial or ‘outside the lab’, making it difficult to know whether they will genuinely work in the real world. Traditional clinical trials are also not necessarily well suited to the testing of AI systems: they are expensive, slow, and designed to evaluate products that are intended to achieve a relatively narrow and specific effect in a specific population, and to remain ‘static’ after deployment.

AI systems are generally intended to deliver broader benefits (for example, increase the speed and accuracy of breast cancer diagnosis in general) to a much larger population, and are both dynamic, developing as new patterns are ‘learned’, and lacking in resilience to external changes (for example a change in the demographic make-up of a hospital population).

Third, and finally, there are ethical considerations related to the risk that, unless carefully managed, the implementation of AI might undermine the NHS’s core values: patient centredness, for all, based on need, not ability to pay. Limitations of the empowerment argument (information does not automatically translate into action) may, for example, undermine individual patient autonomy. Issues such as bias might mean that greater reliance on AI significantly increases health care inequalities. Encouraging self-surveillance and over-medicalising all aspects of life may make ‘health’ feel like an unattainable goal and leave individuals feeling permanently ‘sick’ or in need of constant improvement, creating psychological harm and putting undue pressure on front-line services.

Not getting ahead of ourselves

None of these considerations mean that the new government should stop investing in AI’s potential to help the NHS overcome its current challenges. However, it is imperative that AI is not perceived as a ‘silver bullet’ to improve productivity and it is not deployed before the necessary foundations have been laid.

This means that the Department of Health and Social Care, in collaboration with NHS England and other arm’s-length bodies, must develop policy that targets the entire AI development pipeline. This includes everything from electronic health record design and the creation of well-curated, deep, detailed and representative datasets, to medical device and liability law, health technology and national screening committee assessments of AI, and comprehensive post-market surveillance. It is only by taking this systemic and pragmatic approach that NHS AI implementation will be helpful not hurtful.

Dr Jessica Morley is a postdoctoral researcher at Yale University’s Digital Ethics Center. She previously worked for the Department of Health and Social Care, and latterly NHSX, developing NHS digital, data and AI policy.