Healthcare has come a long way thanks to human endeavour — but even occasional human errors can have disastrous consequences. A misdiagnosis can mean undergoing unnecessary treatment; a misread scan can mean a patient is wrongly given the all-clear. Potentially life-altering or even life-ending mistakes can be made, even by excellent doctors. So the news earlier this year that computer programs can detect breast tumours better than doctors was hailed as ‘a huge advance’ by researchers. The study, published in the journal Nature, compared the results of mammograms analysed by doctors with the same images read by a machine that had been ‘taught’ to identify tumours.
When the computer was asked to read images from nearly 29,000 women, the number of missed cancers — known as false negatives — fell by 2.7 per cent compared with when a single doctor reviewed the scans, while the number of mammograms incorrectly diagnosed as abnormal (known as false positives) decreased by 1.2 per cent. The machine was as good as two doctors working together — the current system for reviewing mammograms.
With thousands of women a year misdiagnosed in the UK, any improvement in the reading of scans would be welcome — and with an estimated shortage of more than 1,000 radiologists (doctors who interpret scans), using technology to do the work of two doctors could free up much-needed time for other tasks. ‘This is a huge advance in the potential for early cancer detection,’ said study author Dr Mozziyar Etemadi, an assistant professor of anaesthesiology at Northwestern University in Chicago. However, more research is needed to work out how such a system could be introduced, he added. Artificial intelligence (AI) like this — advanced computer software which not only carries out tasks but ‘learns’ from the results — is hailed by some as a panacea for the NHS, which is strained by increasing patient numbers and stalling recruitment and retention of doctors and nurses. Last year, the Government announced a £250million AI laboratory that will bring together research to find solutions to common healthcare challenges. ‘The idea is that AI can take some of the workload so that doctors can spend more time with patients,’ says Sarah Deeny, assistant director of data analytics at the independent charity The Health Foundation. It does look promising, and not just in the field of breast cancer diagnosis. Research shows AI could revolutionise diagnostic testing, predict the most effective treatments and improve how hospitals are run.
Another project by the UK tech company Brainomix, backed by pharmaceutical giant Boehringer Ingelheim, uses AI to interpret brain scans of people with a suspected stroke. The Brainomix program analyses CT brain scans in one minute (stock image)
For example, it is helping NHS Blood and Transplant predict how much blood hospitals will need on any given day, resulting in 50 per cent less waste. A trial, published last year in Nature, showed AI was better than specialist doctors at spotting lung cancer; it also boosted detection of the cancer by 5 per cent, while cutting the number of people falsely diagnosed by 11 per cent. It can also identify skin cancers with the same accuracy as doctors, and was as good as humans at diagnosing more than 50 eye conditions in another study.
AI has also been developed to diagnose atrial fibrillation — an irregular heartbeat. And last year, University College Hospital in London came up with an algorithm to flag up patients most likely to skip appointments. Using records from 22,000 MRI scan appointments, the program identified 90 per cent of patients who would not attend, who could then be targeted with reminders. Almost eight million appointments were missed in 2017/18, according to NHS figures, each costing the NHS around £120. Addressing the problem could save almost £1 billion — equivalent to 257,000 hip replacements. Another project by the UK tech company Brainomix, backed by pharmaceutical giant Boehringer Ingelheim, uses AI to interpret brain scans of people with a suspected stroke — a blockage in the blood supply to the brain.
By correctly identifying where the blockage is and the extent of the damage, patients can be given clot-busting drugs within the crucial four-and-a-half hour timeframe, or surgery to restore blood flow to the brain, to give them the best chance of recovery. Currently, reading by doctors of these scans is ‘inconsistent’, according to Dr George Harston, the company’s chief medical officer. ‘Problems do get missed, particularly if a less experienced doctor is looking at the scan or it is a non-specialist hospital.’ The Brainomix program analyses CT brain scans in one minute. This stroke technology has already been adopted in 20 NHS hospitals, but in most cases, AI is not yet widely used in the day-to-day running of the NHS.
‘We are beginning to see the benefits of AI in some healthcare contexts, such as reading scans,’ says Dr Rebecca Rosen, a GP in south London and fellow at the think tank the Nuffield Trust. The key thing which makes artificial intelligence just that — intelligent — is it can learn from the information it receives. But this creates its own challenges. ‘AI is driven by the data that is fed into it from which it learns,’ says Hugh Whittall, director of another think tank, the Nuffield Council on Bioethics, which specialises in the ethical and social aspects of health policy.
Last year, University College Hospital in London (pictured) came up with an algorithm to flag up patients most likely to skip appointments
‘If you only have access to data from a certain population, is that serving all patients? We know from facial recognition programs, for example, that AI is not as good at recognising the faces of darker-skinned people, probably because they are under-represented in the data used to develop the program, and that bias is reinforced by the learning the AI does, making errors more likely.’ There are also concerns about protecting the patient data needed to ‘train’ these robots.
In 2017, the Royal Free Hospital in London was criticised for sharing 1.6 million patient data records — including information on mental health history and abortions — with Google’s AI division, DeepMind, for a trial to monitor and diagnose acute kidney damage. The Information Commission ruled the hospital had not done enough to protect patients’ data. So, is there a danger that we are solving one problem with AI, but creating another? After all, AI is costly and makes mistakes, too.
‘Because of the human nature of healthcare, patients want conversations with real doctors, so AI is most likely to be part of the technology that helps with diagnosis and supports decision-making rather than taking over diagnosis,’ says Mr Whittall. Sarah Deeny agrees, saying it is most likely to be helpful for ‘less high-risk, backroom tasks’ such as making best use of operating theatre time so theatres don’t stand idle, or filing ‘normal’ blood test results in patient records.
‘You don’t want it to result in unnecessary extra tests and clinician time,’ she says. ‘For example, some smartphones can now carry out pulse checks for an irregular heartbeat. However, if this technology were used by the public, healthy people could be told they have an irregular pulse when they don’t. This could result in unnecessary additional testing or treatment, and anxiety for patients.
‘Will a highly anxious, risk-averse patient — or, indeed, any patient — ever be reassured by a computer diagnosis? The role of the doctor is not just to diagnose but to reassure and [guide] patients through their diagnosis and treatment.’ But it’s important not to overlook alternative ways of solving health problems — possibly even more effectively, adds Mr Whittall. ‘There can be a tendency with new technology to find things for it to solve,’ he says. ‘Take malaria, for example. In our search for drugs and vaccines, there is a danger we overlook bed nets which can prevent malaria. ‘There are still many unknowns about AI, so it is important that we don’t push this technology through too quickly.’