Eye scans from 150,000 National Health Service (NHS) patients in the UK will be used to test commercial artificial intelligence tools that could be rolled out to spot the warning signs of diabetic sight loss. But researchers aim to avoid a repeat of previous NHS data-sharing scandals by ensuring that records are anonymised and that AI tests are only run on servers owned by NHS trusts.
Two million people with diabetes in the UK have an annual scan to check for signs of retinopathy, which is the leading cause of blindness in the UK鈥檚 working age population. Each check involves one or more photos being taken of each eye, depending on potential signs of illness, and screening of these 10 million or so images a year is carried out manually by ophthalmologists, optometrists and other trained technicians.
Last year, at St George鈥檚, University of London and her colleagues ran a trial in which AI systems checked scans from 30,000 patients in London and Gloucester. The AI reduced the number of manual checks needed by 50 per cent, and had no false negatives. Now Rudnicka鈥檚 team is working on a larger study, which she calls a 鈥渞eal game changer鈥�. Several companies will be invited to use NHS data to assess their AI tools for potential future roll-out across the UK.
Advertisement
鈥淭he idea behind screening is to prevent sight loss, and it鈥檚 been shown to be very effective. But it鈥檚 just that it鈥檚 a colossal amount of work,鈥� says Rudnicka. 鈥淐utting down the workload by half is massive.鈥�
The team鈥檚 previous study collected data mostly from white people, so the upcoming trial aims to collect more data from people of other ethnicities to ensure that any systems developed have no racial bias. Because the back of the retina has the same pigment as the rest of the skin, scans from different ethnicities can appear very differently, and AI trained on images from predominantly white subjects could perform badly on scans from people with other skin tones.
Some 150,000 people who go through screening under St George鈥檚 University Hospitals NHS Foundation Trust, University College London Hospitals NHS Foundation Trust and Moorfields Eye Hospital NHS Foundation Trust will be included in the trial scheme. Rudnicka told New 杏吧原创 that an opt-out could be looked into if requested by patients, but it wasn鈥檛 clear if they would be routinely told about the trials during screening.
Previous trials of AI in the NHS have proved controversial. UK AI company DeepMind, which is owned by Google parent company Alphabet, announced in 2016 that it was working with the NHS to predict or diagnose acute kidney injury. But a New 杏吧原创 investigation revealed that a data-sharing agreement with the Royal Free London NHS Foundation Trust would give DeepMind access to comprehensive data on 1.6 million people, including sensitive information such as whether they had been diagnosed with HIV or had ever had an abortion. That agreement was later found to break data protection laws.
This new trial will anonymise data by removing personal information, storing the resultant data set on servers inside the NHS. Companies will be able to install their code on these servers and test it using the data, but Rudnicka says that they will not be able to remove any part of the data set. The AI will only assess images for warning signs and cannot learn from the data set. The final list of companies taking part hasn鈥檛 yet been decided but there may be as many as half a dozen.
Sam Smith, who runs health data privacy group MedConfidential in the UK, would like to see assurances about who benefits from the results of the trial and that patients will be informed that their data is being used. 鈥淲ill everything they find be published to advance knowledge rather than benefit private interests,鈥� he says. 鈥淥r will patients have their data fed to corporate AI without any choice and without patients even being told?鈥�
