Training fairer medical AI needs diverse data, but hospitals are restricted in what they can share for privacy reasons. Here, I will discuss our new, easy-to-deploy way for hospitals to take part in AI development without sharing data, and our learnings from a pilot deployment across 4 NHS Trusts. Federated learning (FL) was first developed by researchers at Google as a way to train AI models without moving data. Researchers at NVIDIA, Rhino Federated Computing and University of Pennsylvania have since deployed FL in to hospitals to develop clinical models, but deployment relied on specialist technical expertise at every hospital taking part. Using cheap micro-computers, we built a platform for any hospital to easily take part in training and testing AI models without needing to share patient data. We developed software for FL and loaded it on to Raspberry Pi 4B devices, delivering ‘ready to go’ federated clients to hospitals. Using our approach, four NHS hospital groups developed and evaluated a COVID-19 screening test while retaining full custody of their data throughout, together building a more performant model. By making it easier to train models without moving data, we hope our new full-stack federated learning approach may lead to better and fairer models, while respecting patient privacy and data sovereignty.
Bio:
Clinical academic at Oxford, Profile: www.oncology.ox.ac.uk/team/andrew-soltan
Paper: www.thelancet.com/journals/landig/article/PIIS2589-7500(2300226-1/fulltext
Podcast: open.spotify.com/episode/4KfI1GUjS3nzlYKoW8AgVs
