Researchers have developed a new, easy-to-use technique for hospitals to contribute to the development of artificial intelligence (AI) models, without patient data leaving the hospital’s premises.
The technique which builds on recent advances in decentralised machine learning, uses inexpensive pre-programmed micro-computers, making it easy to deploy in hospitals and cheap to scale up.
Due to the need to safeguard patient privacy, hospitals are often limited in the data they can share to support the development of AI algorithms. Federated learning was developed as a way to train AI algorithms without moving data. Researchers around the world have been working with major technology companies to study how it can be applied in healthcare systems, including in the NHS. However, there has been limited uptake of federated learning in hospitals, in part because its deployment often needs specialist expertise at each hospital taking part in the AI development.
So, the researchers developed and piloted a new technique, called full-stack federated learning, where the software is pre-bundled with inexpensive microcomputing hardware, to make a ‘plug-and-play’ system that is easy for hospitals to deploy. Hospitals taking part in the AI development are sent ‘ready to use’ devices that can be quickly set up, without needing an expert in federated learning onsite.
David Eyre, Professor of Infectious Diseases at the Big Data Institute and a consultant at Oxford University Hospitals NHS Foundation Trust, is a co-author on the study which is published today in The Lancet Digital Health. He said ‘This is a really exciting development that could see hospitals from across the NHS working together to build AI tools to improve patient care, without any patient data ever having to leave each hospital. This tool could have really made a difference during COVID, and we are aiming to have tools like it in use in the NHS in the near future.’
The researchers hope that the new scalable approach, developed with support from the National Institute for Health and Care Research (NIHR) Oxford Biomedical Research Centre, could help address some of the privacy concerns around training AI models using patient data, and lead to models that are more representative and perform more fairly.
The study was led by Dr Andrew Soltan, NIHR Academic Clinical Fellow at the University of Oxford, and Oncology Specialty Registrar and Fellow in Clinical Artificial Intelligence at OUH. He said ‘Building AI models that are fair and inclusive is most achievable when models can be trained with diverse data. Sometimes that might mean training models using data from different parts of the UK, and sometimes also working with data from abroad.
‘Rather than asking hospitals to shoulder the technical burden of taking part in the federated learning, our new approach meant that we, as researchers, did as much of the set-up as possible ourselves before the devices reached the hospitals. By making it easy to train models without moving data, we hope our new full-stack federated learning approach can lead to better and fairer models, which can be developed more quickly, while respecting patient privacy and data sovereignty laws. Together, this might pave the way for the kinds of AI interventions that will lead to improvements for patients in the NHS and those abroad.’
The team demonstrated the new technique by deploying Raspberry Pi 4 micro-computers, at a cost of £45-85 per hospital, to train and validate a screening test for COVID-19 in emergency departments, without the patient data ever leaving the hospitals’ premises.
The micro-computers were preloaded with the software needed to perform federated learning and sent to the NHS trusts. Staff at the hospitals could set the devices up quickly and took part in federated training and calibration to predict COVID-19 status using clinical data from pre-pandemic admissions and COVID-19-positive cases from the first wave. This allowed the researchers to develop a cross-site (global) AI model, which was then evaluated for admissions to three of the NHS Trusts during the second wave.
The results showed that federated techniques improved the performance of the AI model by 27.6%, when compared to the performance of models trained using just an individual hospital’s data, and that the federated model generalised well across sites, making it safe and effective to use.
The Oxford team previously developed an AI test to rapidly screen patients arriving in emergency departments for COVID-19 using clinical information routinely available within the first hour of coming to hospital, but in previous work had centralised the data at the University of Oxford.
The four hospital groups taking part in the pilot were Bedfordshire Hospitals NHS Foundation Trust, University Hospitals Birmingham NHS Foundation Trust, Oxford University Hospitals NHS Foundation Trust, and Portsmouth Hospitals University NHS Trust.