A Survey on Federated Learning for Privacy-Preserving Healthcare: Towards Differentially Private Early Sepsis Prediction

Abstract: Hospitals hold huge amounts of sensitive patient data, but privacy rules and data-ownership concerns mean this data usually cannot be pooled to train machine learning models. Federated Learning (FL) gets around this by sending the model to the data instead of the data to the model. Each institution trains on its own records and shares only model updates. In this paper we survey nineteen studies on federated learning in healthcare published between 2022 and 2026, in ascending order of their citation numbers, which follows the order in which they are first mentioned in the Introduction. The studies fall into seven groups: reviews of FL in healthcare, privacy attacks on FL, cross-hospital generalisation of sepsis models, the dataset source, the base paper on federated sepsis early warning, related work on federated sepsis and Intensive Care Unit (ICU) risk prediction, and differential privacy in FL. For every study we note the objective, the method, the main results and the gap it leaves. The picture is mixed. Federated models can beat single-site models in multi-hospital ICU risk prediction, yet only 5.2% of 612 reviewed healthcare FL studies report real clinical use and only 27.7% include an explicit privacy mechanism. Sepsis models can also lose accuracy at new hospitals, and attacks on FL can reveal not only whether a record was used but which client held it. The ten sepsis studies handle privacy very differently from one another: five report no privacy mechanism in the available text, two run clipping or noise tests with no formal guarantee, one tests a single privacy level, one is centralised only, and one never states its privacy budget. As far as we could tell from the available texts, none gives a patient-level privacy guarantee with accounting across federated rounds. We end by listing the open challenges and by placing FedCare, our planned study of differentially private federated early sepsis prediction across two hospital systems in the open PhysioNet/Computing in Cardiology (CinC) 2019 data, against this gap.

Keywords: Federated Learning (FL), Healthcare, Early Sepsis Prediction, Differential Privacy (DP), Differentially Private Stochastic Gradient Descent (DP-SGD), Membership Inference, Non-Independent and Identically Distributed (Non-IID) Data, PhysioNet/Computing in Cardiology (CinC) 2019.