“Part of my PhD research was monitoring the data that came off the engines of Airbus A380, Boeing 787 ‘Dreamliner’ and the Eurofighter Typhoon,” says David Clifton, Professor of Clinical Machine Learning and Head of the Computational Health Informatics Lab (CHI) at the Oxford University. “It was great fun. I used the 900 engines’ worth of information to create patented systems that would monitor the condition of other jet engines and sold them to Rolls Royce.”
Professor Clifton thought he could repeat his success when he began his post-doctoral training at the emergency department of the Oxford University Hospital. He was wrong. “Back then, patients had to wait about four hours before they were attended to,” he added. “I was not able to do anything because there was no data. Whatever we had about the patients was all in papers.”
Things changed over the years as healthcare systems were digitized, but nonetheless, Professor Clifton believes more can be done to make the most of healthcare data that are locked behind electronic health records (EHRs). That’s why CHI pledges to design data-driven interventions and put them in the hands of clinicians. “A lot of the work that CHI does is consultancy,” Professor Clifton explains. “We are based on the medical campus and 30 seconds away from our clinical colleagues so that we can sit down with them, understand their challenges and try to work out what engineering techniques we have to overcome the challenges.”
Specifically, CHI is capitalizing on machine learning, which is known for processing huge quantities of data all the time. The team had created three predictive models based on patients’ physiology measured at the start and progressively as they were admitted. These models not only forecast the conditions of patients. They are also recommendation providers as clinicians can compare the physiology of new patients to the physiological patterns of admitted patients and take the most appropriate actions.
CHI had also designed a deep learning neural network to find out which tuberculosis mutations are resisting to present treatment. Traditionally, researchers must grow the bacteria that cause tuberculosis in the lab and observe the drugs that deplete it. The process could take up to six weeks as the growing time of tuberculosis bacteria is so long. With machine learning, CHI sequenced and gathered thousands of tuberculosis genomes and their potential mutations, to sift out those that are drug-resistant. All these can be done within a single day, without the need for a microbiology lab.
Given the ongoing COVID-19 pandemic, CHI focused on coming up with a better testing procedure. “In addition to helping the emergency department to predict bed usage and severity of patients, I think a big bottleneck, at least in the UK, has been testing,” Professor Clifton says. “A friend of mine took two days off from his clinical duties waiting for test results, which eventually came back as negative. He could have been working during those two days treating patients.”
Indeed, as Professor Clifton points out, existing COVID-19 tests take 24 to 36 hours to return a result. They have limited sensitivity as only 60-70% of the infected patients will be identified. Besides, the tests require scraping cells from the back of the throat which often puts testers at risk. The AI-based test was co-developed by the CHI team and the emergency department of Oxford University Hospital using routinely acquired blood tests. It’s real-time, highly sensitive (identify more than 80% of infected individuals), effectively cost-free and needing only a regular blood sample which can be taken by GPs and for all hospital patients.
“My friend working at Microsoft once joked, ‘when you are trying to recruit researchers, you’d tell them you are doing machine learning in healthcare. When you are trying to sell it, you’d call it AI in healthcare,” Professor Clifton says. “In reality, I am building a track record of AI in infectious disease, AI in population health, AI in the emergency department, and the list goes on.”
CHI opened its second site in Suzhou, China in 2018, with support from the Chinese government so that researchers can harness more data to build AI-driven tools that have a wider impact.
In 2019, the Wellcome Trust funded its first flagship center consisting of a CHI and Oxford University clinical research unit in Vietnam to work with the local infectious disease facility. “We want to make AI available to middle and low-income countries, where they have limited access to resources and medical experts.” Professor Clifton adds. “AI-driven tools were built using everyone’s data, so they should be used on as many people as possible too.”