Federated AI – Accurate and privacy-preserving learning from distributed medical data
Medicine is undergoing a data revolution, with AI being the engine of change. To unlock this potential, AI algorithms need to learn from very large datasets scattered across multiple hospitals and multiple countries, all in a privacy-preserving and transparent manner. New approaches to algorithmic learning based on recently developed concepts of Federated Learning and Differential Privacy provide the mathematical framework to enable this vision. This research project will develop a new set of algorithms and associated software platform to enable federated-learning at scale, in a way that is respectful of the hospital IT infrastructure and associated privacy needs, thus ensuring the trust of our patients. The project will tackle theoretical problems such as the maximisation of privacy while maintaining model accuracy, and applied problems such as how to learn across multiple hospitals with non-matching care pathways (i.e. multi-task).
Stroke data from KCH (KERRI ethics) and UCLH (High-Dimensional Neurology programme ethics)
Federated-Learning, Artificial Intelligence, Multi-task Learning