AI and A&E
Why it’s easier to forecast the weather than hospital admissions
The AI revolution is sweeping through healthcare, just like every other industry, with claims that smart machines will soon do everything from discovering new drugs to detecting cancer in a scan more accurately and efficiently than any human.
So far, there has been lots of excited talk, far less real world implementation of artificial intelligence. But this week came news of a project that promises to cut NHS waiting lists by using AI to forecast how many people will be admitted to A&E in hospitals in England.
A new tool, developed by the NHS in conjunction with the AI company Faculty, will provide 100 hospital trusts with data on expected A&E admissions up to three weeks in advance. This will help hospital managers understand when demand for emergency treatment is likely to peak or fall, enabling them to make better decisions about staff deployment and other logistical challenges.
One key benefit could be allowing staff to be redeployed to tackle the backlog in non-urgent operations on days when demand for emergency care is expected to be low.
While this use of AI is not as eye-catching as robot surgeons or diagnostic chatbots, it could provide more practical benefits if it means scarce NHS resources can be used more efficiently and thereby cuts waiting lists.
So how does it work? The NHS briefing talks of “sophisticated modelling techniques” and “innovative machine learning technology” - which boils down to chewing through all sorts of data and sprinkling some algorithmic magic on top to get a result. At the beginning of the pandemic Faculty was drafted in to work on forecasting Covid admissions and this new tool builds on that project.
The data includes information about activity levels in previous years, the number of local Covid-19 cases and vaccination rates and public holidays - and will be expanded to include weather conditions.
That doesn’t sound hugely sophisticated and I wondered how the A&E prediction tool compared to the existing system for forecasting activity levels. One might imagine that even before the pandemic hospital managers were looking ahead and trying to predict how busy A&E might be a week on Monday.
It appears that in years gone by there were a few spreadsheets flying around, but there was no systematic retention of any forecasts, so it is impossible to say how much better the AI tool performs - though it appears the hospitals which took part in a pilot were very happy with the accuracy of its predictions.
It all brings us back to the age-old question of how we define artificial intelligence and whether the current hype around it is really justified. I’ve been musing on the comparison between the kind of innovation developed by the NHS and Faculty and the technology behind weather forecasting.
That too involves getting computers to crunch vast amounts of data and then make predictions. The BBC weather app tells me, for instance, that a week today it will be 14C with light cloud in my neighbourhood. Yet if that proves correct we will not be impressed or celebrate it as another triumph for AI, while if instead it is 4C and sleet is falling out of a slate grey sky we will jeer at another failure for the forecasters.
Now I assumed that this was a vastly more complicated task than predicting how busy the emergency department of a hospital was likely to be. Then I spoke to one of the UK’s leading experts on artificial intelligence, Professor Michael Wooldridge of Oxford University, and he put me right.
He explained that weather forecasting, which involves breaking the globe down into cells with current readings of air pressure, temperature and so on and then modelling what will happen next, had got much more accurate as computing power and the quality of the data improved.
But it was still more straightforward to forecast that it will rain next Monday than to work out how many people will turn up at A&E:”It's really fairly well understood laws of physics, there's nothing deep about the mathematics there, it's just the scale of the modelling challenge.”
Forecasting hospital admissions, on the other hand, involved something far less predictable, human behaviour. That is where machine learning comes in, identifying patterns in the data and then making predictions.
If this tool is as accurate as has been claimed then hospitals around the world will be interested in it and it could have some commercial value. I had assumed that Faculty would profit from that but it seems the intellectual property belongs to the NHS.
Some have expressed concern about the private sector’s role in big data and AI projects for the NHS, suspecting that the health service is being naive about the relationship, giving easy access to its crown jewels. But, try as I might, I struggle to see a downside to this partnership which could help to cut waiting-lists while proving a nice little earner for the NHS.
I'm afraid there is no AI that can predict A&E admissions a day in advance, never mind a week. This is typical over-hyping of technology for the benefit of nobody but the investors in companies like Faculty. There are already ways to make statistical forecast for admissions, and these tools use machine learning to refine them a little.
Having dated a couple of nurses over the years, they were not bad at predicting whether they would be busy or not. Certain things were certainties - Friday night would see an increase in fights and alcohol related issues, Monday mornings always had several people coming in because they couldn't get to see their doctor, and so forth. Bad weather would see an increase in cuts and bruises and more serious accidents. Snow would see more car accidents but less bike accidents as bikers didn't use their bikes in the snow.
So, all these factors, which I assume are used with others in the AI version of the nurse's experience, fed into the system.
But for this system to work (just like with the old system), you also need the ideal number of staff available somewhere. You can't drag just anyone into A&E in the same way that you can't just deploy A&E staff elsewhere. There has to be specialist training and experience to match where you are sent. And if there are not surplus staff, which I assume is an efficiency they are looking for, how does the AI A&E modelling take into account sudden surges elsewhere in the hospital?
Some surges, for instance, will involve both. A terrible, large-scale accident not only puts pressure on A&E, but down the line into theatres and wards. And those are unpredictable events, of course.
Although I can see how this can work to an extent, I worry that it is probably too limited because it doesn't apply to the hospital as a whole, and also, is missing one vital bit of data - staff gut instincts based on their holistic experience. Gut instinct is an incredibly important human faculty that has saved people countless times over thousands of years, especially when it comes to human behaviour.
A last thought: this is beginning to sound like psychohistory, Asimov's fictional science developed for the Foundation series. That was a long-term, predictive science of human behaviour taken to the extreme of whole countries and empires. But even in that, there was a problem - The Mule: The unpredictable creature created from two similar creatures. The science knew a mule could happen, but it couldn't predict it. And it brought down the empire.