How to bridge the NHS’s gulf on AI adoption

Sums of cash greater than the GDP of many nations are being thrown at AI for health, yet few applications have scaled our hospital walls into the promised land of routine clinical and operational use. We hear AI can diagnose chronic obstructive pulmonary disease from a cough, meanwhile in the NHS, lung function testing capacity is gasping from over-exertion. With just a photo of a person’s tongue or a retina, AI models can detect a dizzying array of diseases instantly.

But getting a quick GP appointment is easier said than done for many patients. Despite the jaw-dropping capabilities AI tools are now acquiring and the burgeoning patient needs, we know that most AI innovations will never make it through the hospital’s doors. Why?

The reason is implementation, and such tech adoption challenges are far from new.

Take the electric lightbulb – revolutionary tech in its day. It was safer and brighter than candle or oil, and promised to transform dark winters and nights beyond recognition. The man often credited with this invention is Thomas Edison. But, in fact, he didn’t invent the electric light. A Brit, Sir Humphrey Davy, had made an electric lamp over 70 years earlier. 

Edison’s genius was not as an inventor. His talent lay in cracking the implementation challenge and making the tech work outside of the lab. He made bulbs that lasted longer, he made them cheaper, and he built the first grids to get electricity into people’s homes. He understood what was needed to deliver adoption and spread of new technology, and that was what changed the world.

If Edison were here today, he would say the real world focus is missing with AI in health. He would ask how we make it reliable and safe in a hospital setting? On adoption, he would say to invest in the technical and human infrastructure needed to plumb AI in. And he would also ask how we bring people with us, when it radically changes and maybe even threatens their jobs?

Many years ago, I remember discussing with a radiologist whether he would be replaced by AI. The research was starting to show that machine learning tools trained on imaging data were outperforming doctors in the lab setting. Nearly a decade on, he is still in a job, and very few AI imaging tools are in clinical use.

The Royal College of Radiologists has, helpfully, outlined what they think it would take to rollout AI in NHS imaging. It reads less like a turkey’s guide to Christmas, and more like an explanation of why there will be more vegan Christmases to come. The college points out that the infrastructure is not there yet with fragmented, ageing IT systems and painfully slow progress of trusts to cloud computing. They say it needs upfront capital and revenue investment, combined with building in-house AI expertise. 

Despite this negative picture, some trusts are well on the way and will break through. And an intelligent and appropriately resourced national strategy could provide a clear roadmap and remove common barriers being independently tackled across the country.

Take one healthcare AI application, put up in lights at the time of the spring’s capital announcement – the idea of large language models writing discharge letters, along with “Ambient AI” listening into outpatient consultations and writing notes and letters. This focus is understandable, given the vast time spent by doctors writing notes and letters, but it illustrates some of the wider challenges.

Yes, we absolutely need investment in AI, but the unpopular truth is that we will also need the management capacity to transform our inefficient processes as well

The first is Edison’s reliability challenge. The LLMs which write the notes and letters need to be reliably accurate so that the docs don’t spend lots of time proofreading and editing in place of just writing. They need to minimise hallucinations, including clinical diagnoses, which is a current challenge according to research. AI listening in and processing sensitive data requires careful handling, and proving the benefits to patients too. It will come, but will take work.

More fundamentally, we sometimes fail to question the underlying processes and digitise them anyway. Long prose notes and letters have survived almost unscathed in the transition away from paper; instead, maybe, consultants could be told they don’t need to write a clinic letter after every single appointment. Maybe GPs could access clinic notes, discharge summaries, and results directly in their systems, and patients via the NHS App. Maybe, just maybe, we don’t need letters at all?

I have seen many poor processes digitised when the big productivity gains come from asking basic questions about the process itself. So yes, we absolutely need investment in AI, but the unpopular truth is that we will also need the management capacity to transform our inefficient processes as well.

Two shining Nobel prizes were won this month by Brits for AI work, a serious reminder of the expertise we have here in the UK, which could be working with the NHS on our greatest challenges. However, unlocking the big benefits will require both AI inspiration and thoughtful implementation grounded in operational reality.