How AI is changing eye care

The importance of interrogating the evidence supporting the use of artificial intelligence tools was highlighted as part of a panel discussion, Optometry 2049: AI and the Future of Vision Care, at Optometry Tomorrow (14–15 June, Harrogate Convention Centre).

Michael Horler, a consultant optometrist to medical technology company, Cascader, told delegates that AI presents a “huge potential benefit” to patients.

“Ophthalmology is the largest service in the NHS, with over 11 million appointments every year,” he said.

He noted that while one in ten NHS outpatient appointments are eye-related, there are substantial waiting times for many patients to be seen.

Horler highlighted the potential for AI to increase the accuracy of referrals to the hospital eye service and support community monitoring services.

“That will then free up space in our NHS for patients who need to be seen urgently,” he said.

Horler noted that there have been some innovative AI products that have failed to gain traction because of poor usability.

“If you’ve got something that requires 14 clicks to make it work, then it’s going to fail, because people just won’t use it,” he observed.

He emphasised the importance of educating practitioners about what factors they should be considering when looking to invest in an AI tool – for example, whether the technology has been independently verified and details about the population it has been trained on.

“People need to be aware of all of these things so that when they are at a trade fair, and there’s a fancy stand with a device, they are able to critically evaluate it and work out whether it is the right thing for them to use in practice,” Horler shared.

Reflecting on where the responsibility lies when decisions are supported by AI, Horler shared that he envisaged AI tools keeping a human in loop – with oversight of decisions recommended by AI.

“Where the responsibility lies is a gnarly question,” he said.

Community optometrist, Aydan Hussain, noted that many of the approved applications for AI within eye care have focused on hospital settings.

“In eye care, even though there is lots of clever work, these tools are not necessarily designed for us to use in primary care. In terms of tangible, day-to-day use – we are still really waiting for this technology to have an effect on us,” he reflected.

Like Horler, Hussain emphasised the value of keeping the clinician in the loop when it comes to AI.

“In primary care, if my signature is on the prescription, I am taking 100% of the responsibility,” he said.

“Having an awareness of the limitations of AI – whether it is black box in nature, how it was trained and how it came to that decision – is important when it is my name on the record card,” Hussain noted.

Professor Alicja Rudnicka, of City St George’s, University of London, shared with delegates that she is involved in research that evaluates algorithms used by commercial companies to identify diabetic eye disease from retinal images.

She noted that there are around 3.5 million appointments within the NHS diabetic eye screening programme each year.

Rudnicka noted that the number of people who are eligible for diabetic eye disease screening is rising in the UK.

“AI presents an opportunity to enable us to triage patients who are low enough risk that they do not require human grading,” she said.

The professor of statistical epidemiology emphasised the importance of evaluating AI tools before they are implemented in the community.

“No algorithm is perfect,” Rudnicka shared.

“What’s really important for these AI tools, is that they are evaluated to make sure they work as intended in the setting they will be deployed in, and that they work equally for all,” she said.

Rudnicka observed that there is a broad spectrum of AI tools – from low-risk tools, such as AI scribes, to higher risk tools that aim to identify specific eye diseases or monitor progression.

She shared that optometrists do not need to suddenly become computer scientists to harness the potential of AI.

“It’s about knowing how to interact with this tool – knowing how it operates and the boundaries within which it operates,” Rudnicka said.

Reflecting on the limitations of AI, Foresight Research chief executive, Dr Wen Hwa Lee, shared an example of an algorithm that was trained to detect skin cancer.

“Because the dataset it was trained on was a mostly Caucasian population, it performed poorly when deployed on non-Caucasian skin,” Lee explained.

Addressing concerns about whether companies are profiting from patient data, Lee noted that any person who has a smartphone is already vulnerable to this.

“You are already being monetised, and for things that will not necessarily serve you,” he said.

Lee shared that when practitioners are discussing data with their colleagues and patients, they should take into account the benefits of sharing data for innovation.

“All the drugs we have today were developed with the help of somebody who very kindly donated their own data. I am up for it – but it is up to you individually,” he said.