How to implement and scale AI solutions safely and effectively

In conversation with Erik Mayer and Adnan Tufail

Can you start by each giving a quick summary of the projects you’ve been leading to transform care using AI?

Adnan: The screening programme for diabetic retinopathy identifies early signs of damage to the eye caused by diabetes, enabling treatment to avoid sight loss. Patients have photos taken of their eyes every year. With over 4 million people with diabetes in the UK, this process generates almost 20 million images per year, which need to be looked at by three levels of trained technician graders. 

We've been working with a team from City St George’s, University of London and Moorfields Eye Hospital to find a suitable AI solution to automate parts of the screening process. This will reduce the cost of screening (by reducing the number of human technicians involved) and speed up diagnosis for patients. 

Algorithms already existed but needed to be independently validated using large scale representative data sets to ensure they work for all patients. Our findings will inform future commissioning, leading to one of the first mass deployments of AI within the NHS. 

Erik: We’ve used AI to analyse results from the NHS Friends and Family Test. At Imperial this generates around 20,000 comments a month about people’s experience of care, but with only a small patient experience team (as is typical across all NHS providers) there just isn’t the capacity to work through that volume of free text data – so much of the feedback wasn’t being used. 

We used natural language processing to pull out key themes and give information back to quality improvement teams. We then looked at ways we could scale this approach across ten other NHS organisations. The code is now freely available to other NHS organisations.

That project led to wider work at Imperial to set up the iCARE secure data environment where we could, in near real time, ingest all the trust’s routinely collected health care data. We now have over 2 billion rows of structured data, but most importantly we capture all the free text data, that's 400 million records, all linked at patient level. 

In essence that’s created a secure AI testbed for us to work out how we can use generative AI in useful ways, and is instrumental in delivery the NIHR Imperial BRC Digital Health Theme objectives. For example, we’re now developing automated discharge summaries, which will contain slightly different information depending on which setting the patient is being discharged to. 

What learning can you share about how to successfully scale use of AI in the NHS?

Adnan: There's this notion that if you deploy AI it will solve everything, but we have to get the basics right. What slowed us down more than anything else was having to interact with extremely stretched NHS IT departments who were understandably worried about things like GDPR. 

That’s a continuing challenge as we scale up for a larger implementation study to test with 200,000 patients – the final step before commissioning. We’ve shown how well these algorithms perform; it's now about having the right NHS digital support to run it in more centres.

Erik: Yes, although our tool wasn't very complicated (you can run it on a laptop), there was always initial fear from ICT staff thinking we needed some great supercomputer and a massive data science resource. Starting to spread to other organisations meant learning to address those instant negative barriers that came up. 

One common thread that always drove success was having exec-level leadership to really own it locally. We didn’t say ‘we've come up with this dashboard, here’s what you need to do with it’. Instead, we said, ‘these are the ingredients to support a recipe, but you've got to slightly tweak and own the seasoning based on your local requirements’. 

How should the NHS be partnering with industry to use AI to its full potential? 

Erik: People are always knocking on the door of the NHS saying, ‘I've got this great tool… can we find a use for it?’ But the question we should be asking is, ‘what’s the problem that we want to solve?’

There's a real opportunity in the relationship between industry, NHS and academia to make this a process of co-design. And for that the NHS needs to be stronger at identifying clear problem statements; this is future of successful life science partnerships. 

Adnan: Yes, it’s helped that our problem space was already well defined, with a very clear metric – to be at least as safe, effective and equitable at screening as the humans currently doing the job. 

Everything else then becomes a competition on price or other metrics of specificity. We could end up with several potential providers. That allows competition and ensures no single company dominates the landscape. 

Even though it's competitive, we’re providing a clear pathway, and I think that’s what companies want. It encourages investment and innovation because people know the rules and how to engage.

How should the NHS be developing and monitoring the use of AI?

Erik: We need to create secure data environments in which we can support the collaboration between the NHS, industry and academia to work together and provide the evidence needed to support regulatory approval and justify commissioning for clear benefits for the NHS. That involves creating AI test beds like we’ve done at Imperial, where you can build trusted relationships. 

Adnan: But we also need to be aware of how vendors’ own single AI model studies can be very different, even when they’re run independently. We don’t currently have the equivalent of drug regulation saying you have to trial things a certain way. 

Erik: Yes, we’re still working out what collating the evidence for regulatory approval looks like.

And how do we monitor it once its deployed? AI can use models that continue to learn as they're exposed to new data. That really changes what ongoing regulatory processes need to look like. 

If I launch something new, how do I know it’s going to continue to behave the same way next month? What does that data-driven, evidence-based check look like? It's almost like a learning health system really.

Adnan: The current models being tested have fixed operating points that do not alter how they behave over time. We have wrestled with this problem for the future. A potential safe way forward is that you run a parallel model to that used to triage patients, that continues to learn on the data, but you don't alter the patient pathway using this continually learning part of that model. Then you fix the operating point again that would require another validation study later… iterative changes that you can regulate.

Erik: But organisations won’t be running one algorithm, they’ll be running thousands across direct care, operational efficiency, and back-office functions.

The other way companies are trying to address this is by identifying the fail safes. The points where we know tools are starting to misbehave. It’s likely we’ll have to use AI itself to signpost when things go wrong. A degree of automation for evaluation will have to be involved, otherwise it's not sustainable. 

What other challenges does the NHS face in coping with how fast AI technology is developing? 

Erik: Tools like digital scribes are really shifting the dial – and that’s where an awareness of human behaviour in all of this is critical.

I might have a long, complicated discussion with a patient about an operation, and AI is capturing the entire narrative. That’s four or five pages. Will I have the time to check it all in detail?

I also suspect if you know there's going to be a full description recorded, that consultation will go on longer and you’ll maybe answer the patient's questions differently. That's human behaviour. 

It just shows why we have to consider human factors and introduce AI in conjunction with the health care workforce. I don't think that’s currently happening as much as it probably needs to.

Adnan: It comes back to your point about developing clear problem statements. Both of our projects had that from the start, and we haven't deviated and got carried away with all the other things we could be doing. Instead, we’ve focused on applying the right technology to address the problem in the simplest and safest way. 

Erik: Indeed. The technologies will continue to advance, but you can't boil the ocean tomorrow. It's think big, start small – that's the way we've got to keep steering this ship, and most importantly with our patients and population at the controls.

This content originally featured in our Deep Dive newsletter, which explores perspectives and expert opinion on a specific health or health care topic.