The case for scaling AI in screening for disease

We must understand what responsible AI adoption requires when moved out of controlled trials and into the realities of the NHS, writes Professor Alicja Rudnicka, professor of health and medical sciences at City St George’s, University of London

The current direction from government points toward a growing role for AI in diagnostics and screening. Recent announcements have emphasised its potential to increase capacity, reduce variation and support earlier detection of disease. And there are areas of care, particularly imaging-based services, where AI could make a measurable difference.

However, introducing AI into the day-to-day delivery of healthcare requires more than demonstrating technical accuracy or producing promising pilot results. It requires clarity on digital infrastructure, workflow integration, cost, fairness and confidence from both clinicians and patients alike.

It also depends on the embedded solutions that are already underpinning service delivery, such as national screening management systems and referral platforms, which need to interact seamlessly with any AI solutions selected for scaling.

Before AI can move from policy papers to real NHS pathways at scale, we need practical examples of how these systems behave when tested under the same pressures faced by clinicians every day.

It’s a gap that research has begun to fill, including the publication of our recent study in The Lancet. Led by my team at City St George’s, University of London it presents findings from the world’s first large scale evaluation of commercial AI systems for diabetic eye screening using real world data.

We need practical examples of how AI systems behave when tested under the same pressures faced by clinicians every day

We reviewed more than 1 million retinal images and found that AI can detect sight threatening disease with 96-99%  accuracy and process images in seconds rather than minutes. It demonstrates the potential of AI and the importance of understanding how these tools behave once introduced into real service environments.

Why evaluation matters

The research has helped us to understand the practical dimension in more depth – by testing tools against existing NHS grading pathways using real images, we can observe how they perform when exposed to the variation and complexity that services manage every day.

Not to declare winners or to validate specific products, but to understand what responsible AI adoption requires when moved out of controlled trials and into the realities of NHS data and workflows.

Several themes emerged that are critical for clinical use.

Firstly, that independent evaluation is essential. Vendors naturally want to showcase their performance, but the NHS needs assurance that testing is impartial, uses representative data and doesn’t rely on small or narrow cohorts.

Secondly, and equally important, is meaningful digital readiness. Even a high-performing algorithm cannot be adopted if local systems cannot upload images reliably, return outputs quickly or integrate results back into the platforms used by screening teams. Digital infrastructure will determine whether AI is genuinely usable, rather than just technically impressive.

The NHS needs assurance that testing is impartial, uses representative data and doesn’t rely on small or narrow cohorts

Thirdly, efficiency must also be balanced with safety. AI can process images rapidly, but speed only adds value when supported by transparent workflows, clear reporting rules and robust review mechanisms for borderline or complex cases. Clinicians need confidence that AI enhances decision making rather than bypassing it.

Cost is also a critical factor. It must be more cost effective than existing pathways. Those responsible for commissioning and delivering screening services cannot adopt technology that adds cost without improving outcomes or reducing pressure on staff.

Lessons for adoption

Our hope is that this work offers practical steps that healthcare teams can prioritise to scale AI safely and sustainably, whether it be locally or across a national service such as diabetic eye screening.

Lessons for adoption include investing in in-service evaluation. Controlled environments can tell us how algorithms behave under ideal conditions, but real world tests expose the operational variation that truly reveals whether AI is ready for deployment.

Successful adoption of AI relies on supporting rather than replacing the workforce

Another priority is to develop shared digital infrastructure. A central approach that allows screening providers to upload images securely and receive consistent outputs integrated into electronic patient records would reduce duplication and avoid fragmented local implementations.

This would also make it easier for systems like Optomize, which manages most English screening programmes, to interact predictably with AI tools, across both optometry and specialist ophthalmology.

Equally, there needs to be trust and transparency throughout. Staff confidence is essential and a poorly implemented tool could undermine progress across services.

Clear communication, routine monitoring and alignment with evolving Medicines and Healthcare products Regulatory Agency expectations for post market surveillance will be critical to ensuring clinicians feel they can rely on AI safely.

And finally, we must remember that AI isn’t a replacement for clinical expertise. It’s a tool that can free-up capacity, standardise repetitive tasks and allow clinicians to focus on more complex imaging modalities and patient care. Successful adoption relies on supporting rather than replacing the workforce.

A path forward

AI has significant potential to improve the efficiency and consistency of screening services, particularly in the diabetic eye screening programme, which manages more than two million visits each year. Our evaluation demonstrates that when the right safeguards are in place, it can play a meaningful role in supporting the first stage of grading.

The next steps are about translating it into safe and practical use. If we can take this approach, AI can be introduced in a way that strengthens screening pathways and improves outcomes without compromising equity or trust.

This way AI supports clinicians, fits naturally into established services and provides meaningful benefit for the people who rely on these programmes.