Analysis of Deep Federated Learning on Early Prediction of ICU Mortality Risk

Abstract

In the technological age, vast amounts of data are being generated continuously, holding information that could significantly improve various aspects of everyday life, if effectively utilized with machine learning techniques. However, traditional centralized machine learning approaches are subject to privacy and security concerns, which impede the development of robust, generalizable models. Federated learning is a novel approach in the field of machine learning, which addresses the aforementioned issues by allowing multiple parties to collaboratively train a machine learning algorithm without exchanging their local data. One domain that could be significantly benefited from the advancement of federated learning is healthcare, where data from multiple medical institutions cannot be centralized due to patient privacy.

In this thesis, we investigate the application of state-of-the-art federated learning (FL) algorithms on the early prediction of ICU mortality risk. More specifically, we focus on utilizing the rich temporal dynamics of vital signs and laboratory results, in the form of multivariate time series, from the first 24 hours of an ICU stay to estimate the mortality risk in the following 48 hours. Our data originates from multiple hospitals across the US between 2014 and 2015, therefore allowing us to recreate a realistic, multi-center federated environment. In order to gain an insight into the FL framework, with regard to this task, we design a series of experimental scenarios. Our aim is to explore the generalizability and integration of different deep recurrent neural network (RNN) models with FL, the sensitivity of different FL algorithms in the presence of heterogeneous local data distributions and the effect of individual hospitals on the development of a FL model.

We compare our FL models with reference to the centralized machine learning (CML) and local machine learning (LML) approaches. Our results indicate that, in settings with non-IID datasets, the FL models are superior to the privacy-preserving LML models, in terms of AUROC, AUPRC and F1-Score, while they perform slightly worse, in general, than the CML models. Then, depending on the characteristics of the FL participants, in relation to data size and class representation, certain FL algorithms exhibit better suitability for specific scenarios. Moreover, hospitals with both smaller and larger datasets may improve the models' performance on their local data, by participating in FL model training.