One of the country’s biggest hospitals has unveiled sweeping plans to use artificial intelligence to carry out tasks traditionally performed by doctors and nurses, from diagnosing cancer on CT scans to deciding which A&E patients are seen first.
The three-year partnership between University College London Hospitals (UCLH) and the Alan Turing Institute aims to bring the benefits of the machine learning revolution to the NHS on an unprecedented scale.
Prof Bryan Williams, director of research at University College London Hospitals NHS Foundation Trust, said that the move could have a major impact on patient outcomes, drawing parallels with the transformation of the consumer experience by companies such as Amazon and Google.
“It’s going to be a game-changer,” he said. “You can go on your phone and book an airline ticket, decide what movies you’re going to watch or order a pizza … it’s all about AI,” he said. “On the NHS, we’re nowhere near sophisticated enough. We’re still sending letters out, which is extraordinary.”
At the heart of the partnership, in which UCLH is investing a “substantial” but unnamed sum, is the belief that machine learning algorithms can provide new ways of diagnosing disease, identifying people at risk of illness and directing resources. In theory, doctors and nurses could be responsively deployed on wards, like Uber drivers gravitating to locations with the highest demand at certain times of day. But the move will also trigger concerns about privacy, cyber security and the shifting role of health professionals.
The first project will focus on improving the hospital’s accident and emergency department, which like many hospitals is failing to meet government waiting time targets.
“Our performance this year has fallen short of the four-hour wait, which is no reflection on the dedication and commitment of our staff,” said Prof Marcel Levi, UCLH chief executive. “[It’s] an indicator of some of the other things in the entire chain concerning the flow of acute patients in and out the hospital that are wrong.”
In March, just 76.4% of patients needing urgent care were treated within four hours at hospital A&E units in England in March – the lowest proportion since records began in 2010.
Using data taken from thousands of presentations, a machine learning algorithm might indicate, for instance, whether a patient with abdomen pain was likely to be suffering from a severe problem, like intestinal perforation or a systemic infection, and fast-track those patients preventing their condition from becoming critical.
“Machines will never replace doctors, but the use of data, expertise and technology can radically change how we manage our services – for the better,” said Levi.
Another project, already underway, aims to identify patients who are are likely to fail to attend appointments. A consultant neurologist at the hospital, Parashkev Nachev, has used data including factors such as age, address and weather conditions to predict with 85% accuracy whether a patient will turn up for outpatient clinics and MRI scans.
In the next phase, the department will trial interventions, such as sending reminder texts and allocating appointments to maximise chances of attendance.
“We’re going to test how well it goes,” said Williams. “Companies use this stuff to predict human behaviour all the time.”
Other projects include applying machine learning to the analysis of the CT scans of 25,000 former smokers who are being recruited as part of a research project and looking at whether the assessment of cervical smear tests can be automated. “There are people who have to look at those all day to see if it looks normal or abnormal,” said Williams.
Might staff resent ceding certain duties to computers – or even taking instructions from them? Prof Chris Holmes, director for health at the Alan Turing Institute, said the hope is that doctors and nurses will be freed up to spend more time with patients. “We want to take out the more mundane stuff which is purely information driven and allow time for things the human expert is best at,” he said.
When implementing new decision-making tools, the hospital will need to guard against “learned helplessness”, where people become so reliant on automated instructions that they abandon common sense. While an algorithm might be correct 99.9% of the time, according to Holmes, “once in a blue moon it makes a howler”. “You want to quantify the risk of that,” he added.
UCLH is aiming to circumvent privacy concerns that have overshadowed previous collaborations, including that of the Royal Free Hospital in London and Google’s DeepMind, in which the hospital inadvertently shared the health records of 1.6 million identifiable patients. Under the new partnership, algorithms will be trained on the hospital’s own servers to avoid any such breaches and private companies will not be involved, according to Holmes.
“We’re critically aware of patient sensitivity of data governance,” he said. “Any algorithms we develop will be purely in-house.”
Questions also remain about the day-to-day reality of integrating sophisticated AI software with hospital IT systems, which are already criticised for being clunky and outdated. And there will be concerns about whether the move to transfer decision-making powers to algorithms would make hospitals even more vulnerable to cyber attacks. Hospital IT systems were brought to a standstill last year after becoming victim to a global ransomware attack that resulted in operations being cancelled, ambulances being diverted and patient records being unavailable.
Williams acknowledged that adapting NHS IT systems would be a challenge, but added “if this works and we demonstrate we can dramatically change efficiency, the NHS will have to adapt.”
Original article by Hannah Devlin
This post was curated with edits by Gordon Fletcher, Principal Consultant(Engineering & Mobile Technology) at Compumagick Associates can be reached at https://www.compumgickassociates.com/contact, @compumagick