At the beginning of February, Bjarne Pfitzner took part in the International Conference on Artificial Intelligence for Medicine, Health, and Care (AIMHC) in Laguna Hills near Los Angeles - and received the Best Student Paper Award. His work deals with federated learning in the field of digital health - and is titled: "Differentially-Private Federated Learning with Non-IID Data For Surgical Risk Prediction" (Bjarne Pfitzner, Max M. Maurer, Axel Winter, Christoph Riepe, Igor M. Sauer, Robin van de Water and Bert Arnrich).
The PhD student at the Research Group "Digital Health - Connected Healthcare", led by Prof. Bert Arnrich, says about his work:
“Federated learning is a technique for training machine learning models on sensitive and protected data (such as medical data) that was collected at multiple locations and cannot simply be consolidated in a central database.
In our paper, we investigated the interaction of federated learning and so-called differential privacy, the introduction of noise into the machine learning model during the training process. This prevents the risk of reconstructing the training data from the finished model. Specifically, we analyse an application of this methodology in situations with only a few training participants who have small amounts of data that additionally have different statistical characteristics.
We were able to show that in these cases the final model does not work equally well for all participants and therefore does not offer the same added value for everyone. As a potential solution, we show the fine-tuning of the final model to the specific data of the individual participants, which improves the model in all instances.
The paper originates from a collaboration with our partners at Charité - Universitätsmedizin Berlin, who provided the patient data. As a medical use case, we consider the prediction of patient mortality and the need for revision surgery for patients undergoing major visceral surgery.”
