Early ICU Mortality Prediction with Deep Federated Learning: A Real-World Scenario

Abstract

The generation of large amounts of healthcare data has motivated the use of Machine Learning (ML) to train robust models for clinical tasks. However, limitations of local datasets and restrictions on sharing patient data impede the use of traditional ML workflows. Consequently, Federated Learning (FL) has emerged as a potential solution for training ML models among multiple healthcare centers. In this study, we focus on the binary classification task of early ICU mortality prediction using Multivariate Time Series data and a deep neural network architecture. We evaluate the performance of two FL algorithms (FedAvg and FedProx) on this task, utilizing a real world multi-center benchmark database. Our results show that FL models outperform local ML models in a realistic scenario with non-identically distributed data, thus indicating that FL is a promising solution for analogous problems within the healthcare domain. Nevertheless, in this experimental scenario, they do not approximate the ideal performance of a centralized ML model.