AI used to predict Covid-19 patients’ oxygen needs by hospitals including Addenbrooke’s and NVIDIA
Addenbrooke’s and 20 other hospitals around the world teamed up with technology company NVIDIA to use artificial intelligence (AI) to predict the oxygen needs of Covid-19 patients.
Using data from Europea, Asia, North and South America, the resulting tool proved capable of forecasting the amount of oxygen a patient would need within 24 hours of arrival in the emergency department, with a sensitivity of 95 per cent and a specificity of more than 88 per cent.
The researchers used a technique called federated learning, through which an algorithm is trained using multiple decentralised edge devices, meaning that no data needed to be shared with other hospitals or leave its location. All the patient data was also fully anonymised.
Chest X-rays and electronic health information was used to train the algorithm, which took just two weeks. Once it had learned from it, the analysis was brought together to build the AI tool, which can be used anywhere in the world.
The study, called EXAM (for EMR CXR AI Model), is said to be one of the largest, most diverse clinical federated learning studies to date.
Its findings were tested in hospitals across five continents, including at Addenbrooke’s.
“Federated learning has transformative power to bring AI innovation to the clinical workflow,” said Professor Fiona Gilbert, who led the study in Cambridge and is honorary consultant radiologist at Addenbrooke’s Hospital and chair of radiology at the University of Cambridge School of Clinical Medicine.
“Our continued work with EXAM demonstrates that these kinds of global collaborations are repeatable and more efficient, so that we can meet clinicians’ needs to tackle complex health challenges and future epidemics.”
Dr Ittai Dayan, the first author of the study, from Mass General Bingham in the US, where the EXAM algorithm was developed, said: “Usually in AI development, when you create an algorithm on one hospital’s data, it doesn’t work well at any other hospital. By developing the EXAM model using federated learning and objective, multimodal data from different continents, we were able to build a generalizable model that can help frontline physicians worldwide.”
Data from around 10,000 Covid-19 patients was analysed in the study, including 250 who came to Addenbrooke’s in the first wave of the pandemic in March and April 2020.
“Federated learning allowed researchers to collaborate and set a new standard for what we can do globally, using the power of AI,'' said Dr Mona G Flores, global head for medical AI at NVIDIA. “This will advance AI not just for healthcare but across all industries looking to build robust models without sacrificing privacy.”
The National Institute for Health Research (NIHR) Cambridge Biomedical Research Centre (BRC) supported the study, and work on the EXAM model has continued.
Mass General Brigham and the NIHR Cambridge BRC are working with NVIDIA Inception start-up Rhino Health, co-founded by Dr Dayan, to run prospective studies using the model.
Professor Gilbert added: “Creating software to match the performance of our best radiologists is complex, but a truly transformative aspiration. The more we can securely integrate data from different sources using federated learning and collaboration, and have the space needed to innovate, the faster academics can make those transformative goals a reality.”
Sign up for our weekly newsletter and stay up to date with Cambridge science