AI diagnostics predict COVID-19 without testing

UK and US researchers use mathematical model to diagnose coronavirus with 80% accuracy.
12 May 2020 , Katie Coyne

Aritificial intelligence (AI) symptom diagnostics can predict whether someone has COVID-19 with almost the same accuracy as current tests, according to experts.

Research published in Nature Medicine indicates that an AI mathematical diagnostics model can predict – with an accuracy of nearly 80% against current tests – whether someone has the virus based on symptoms recorded over a period of about a week.

A second phase of the research – a collaboration between King’s College London, Massachusetts General Hospital, Nottingham University and science company Zoe – will see a further two trials aimed at producing an accurate diagnosis based on just two days of symptoms. 

Massachusetts General Hospital will run a trial aimed at high risk groups in the Boston area. Those taking part – around 5,000 people – will be tested at the start for COVID-19 antibodies. They will use the app daily and be retested if they display COVID-19 symptoms. A second trial is due to start shortly in the UK.

At the heart of the work is the COVID Symptom Study app, formerly known as the Symptom Tracker. Researchers changed its name to avoid confusion with apps that track and trace. Launched on 24 March this year, the app now has more than 3 million users in the UK. 

Amid all the noise surrounding COVID-19, scientists working with the symptom study have doggedly followed the trail left by the virus, gaining insight into tackling it. So far, it has helped identify COVID-19 hotspots, and proven scientifically, beyond anecdotal reports, new symptoms such as the loss of taste and smell, and provided research in the form of more than 25 scientific papers.

Against a background of difficulty ramping up testing, an AI diagnostic approach could complement and assist. Using symptoms to diagnose in an infectious disease outbreak is not new, but until now has not been possible in practice with COVID-19.

Prof Tim Spector, who leads the COVID-19 symptom study research programme at King’s, said: “If it was chicken pox it would be easy. We just basically say, ‘do you have a rash?’ and that would cover 99% of cases. Okay, you’ve got a rash, stay at home, isolate for a week and don’t come out, and who did you meet recently? Tell them [you’ve got chicken pox]. That’s what the government tried to do initially, and said if you’ve got a fever or a persistent cough then do the same [isolate].”

However, he added: “Because there are at least 14 different symptoms of the disease and they happen at different times over two weeks – and many of them while you’re still infectious before things are obvious – most of these cases got missed. 

“But if you get everyone to log all their symptoms and you combine it with machine learning and millions of people, you can actually work out by mathematical modelling what a likely score is that that combination of symptoms meant, if you had been tested, you would be positive.”

Spector said swab testing was around 70% accurate, and antibody testing around 60-70% accurate for the virus, and the AI model performs around 80% as well as current testing. “There’s no gold standard, so you can’t say it’s better than the test. It’s nearly as good as the test,” he said.

He added the symptom study is very different to the government’s tracking app: “They are tracking them as an individual, and their phone as an individual phone and looking at their contacts. We are not doing anything at the individual level. We can track things back to a rough area and the personal details are secure and not shared with anybody.”

Luminous cubes

CIEH Leadership Programme

Take your career to the next level with the CIEH Leadership Programme.

Find out more