Federated learning and synthetic data across hospitals
Training and validating clinical models across institutions without moving patient data, and generating synthetic perioperative data that can be shared.
A model trained in one hospital has to work in the next one, but patient data cannot simply be pooled. Federated learning trains a shared model while each hospital keeps its data, and synthetic data offers a way to share realistic records without exposing real patients. Each technique is established on its own. Their joint use on real clinical data across institutions is still rarely tested.
The project Secure International Collaboration: Advancing Medical AI with Federated Learning and Synthetic Data (2025–2029) is funded by the Digitalization Commission of the University of Bern. I am a named associated researcher and I conceived and lead its AI and methods component: federated training and evaluation of CARESCORE across Inselspital, Bern and CHUV, Lausanne, and the generation of synthetic perioperative data. The principal investigator of record is Prof. Guido Beldi, with Dr Olga Endrich as co-applicant.
Status: work in progress. No results from this project have been published yet.