AI in Disease Detection


Identifying the contribution of AI

Jess Morley

Jess Morley is a researcher at the University of Oxford. Her work focusses on the use of health data for research and analysis, including the development of AI-based clinical decision support software. She was the co-author of the Government-commissioned Goldacre Review of April 2022. Prior to working full-time in academia, Jess was a civil servant for the Department of Health and Social Care/NHSX.


  • Data permeates the healthcare system
  • The NHS has not always been able to embed large scale technology transformation
  • The NHS is committed to support innovations with potential to save lives
  • AI has the potential to change the way the health service works
  • AI cannot replicate the relationship between patient and clinician.

The Institute for Healthcare Improvement (IHI) Triple Aim for the NHS is to improve the population’s health and improve the experience of care while reducing the per capita cost. This is to be achieved by ‘P4’ medicine; that is, medicine which is predictive, preventative, personalised, and participatory.

Essentially, this means gathering data on an individual all the time, using algorithms to find the level of personal risk and determining personalised risk stratification. Then the appropriate drugs are identified to provide personalised, earlier levels of intervention. This should reduce cost because it is cheaper to treat people early than late – or indeed to prevent people from getting sick in the first place. Ultimately, this should improve the population’s health.

Data at the heart

Data is at the heart of this concept. It permeates the healthcare system the entire time, not only improving our ability to diagnose individual patients earlier, but also improving our ability to record and monitor the overarching performance of the healthcare system and so deliver better outcomes.

Clinical decision support software (CDSS) has existed in the NHS since the 1980s. Further, the first paper promoting the idea that AI might help with diagnosing and diseases was published in 1959, so the concept has been known for a long time. CDSS is not very smart, though; it works on pop-ups based on flowcharts. If I go to my GP, for example, there will probably be a pop-up that appears in the electronic health record system to say: Jess is female, over 25, needs reminding to go to cervical screening. That is not very sophisticated.

More recently, there have been attempts to measure an individual’s hazard ratios compared to different people in the population – i.e. compared to a baseline using large numbers of patient records – and predict the likelihood of adverse reactions to infection. In essence, these are models or algorithms predicting risk, which allow earlier intervention at the point of care.

However, the NHS does not have a good record in large transformations of technology on this scale. The NHS National Programme for IT (NPfIT) was the biggest public sector IT transformation programme in the world. It cost a great deal of money and did not achieve its projected outcomes. However, it did deliver the NHS Spine which allows information to be shared securely across national services, a vital platform for the organisation.

Projects to do with data have also been unsuccessful. For example, the Royal Free and Deep Mind tried to develop an app called Streams, which would alert clinicians to people who were likely to get acute kidney injury. However, they misinterpreted data protection law. If the purpose is direct care, i.e. one doctor talking to another doctor, patient consent is not needed in order to transfer the records. If the purpose is research, patient consent is needed.

The Royal Free thought that because it was developing the app for use in the hospital, this fell under direct care, whereas in fact it fell under research. They had handed over many thousands of patient records without consent and had broken the law.

So should the NHS cease this activity? No, because the NHS constitution states that it is committed to supporting innovation when there is the potential to save lives. And we know that potential is there. But healthcare is complicated.

Healthcare is as much about the dynamic between the patient and the clinician as the treatment itself. AIs cannot replicate that. 

AI is complex, too. Remember, too, that we are not trying to deal with just one type of condition, we are trying to screen every individual for every possible condition at all times. Hardware does not always work, data quality issues can arise. Most of the population is not well-represented in our healthcare data.

The NHS is often presented as having the best healthcare data in the world. And we do, but it needs a lot of work to make it work, then it has to be integrated into clinical systems, you have to pass data protection requirements – and all of these stages have to go well for a project to succeed.


Re-ontologising means fundamentally transforming the healthcare service. AI has the potential to change what counts as knowledge about the body. This is because AI monitors things that we can record quantitatively – heart rate, how many steps a person takes, how much sleep they get. Yet it cannot measure how you feel and what your outcomes are. Often people go to the doctor when they just do not feel like themselves. That is an early indicator that AI cannot measure because it is not quantitative.

Should we only take account of what appears in the data, not what people say about themselves as a person? That would change who has the right to say that they have knowledge about the body – the algorithm or the person. In that case, we would only be considering the ‘data clone’ of an individual. That data clone may not accurately represent the person and their physical body.

We know that healthcare is as much about the dynamic between the patient and the clinician as the treatment itself. AIs cannot replicate that, they are not human. While it is possible to teach a Large Language Model to mimic an empathetic-sounding human, it will never understand what that actually means.

Treating everybody?

Not everybody is – or will become – equally represented in datasets. Not everybody has access to the latest smartphones. People who do not have a fixed address might not appear in electronic health records, although they are equally deserving of care.

Looking to a future where AI helps to prevent disease earlier, we should be focussing on aspects of information and utility. Does this application do something useful? Is it screening for something that we can actually treat? Is it usable – can the clinician actually run it in a clinic and understand what it says? Does it actually work – at the moment, AI is very precise and very accurate, but there is little evidence that it can improve outcomes in the real world. And, then, do people trust it?

If we can address those four things – utility, usability, efficacy and trust – we will have success. If we cannot, we will not.