Federated Learning for Enhanced Performance on Decentralized Healthcare Datasets

Federated learning (FL) has emerged as a new paradigm to enable the application of privacy-preserving machine learning (ML) on decentralized datasets. In this paper, the use of FL has been demonstrated on a medical dataset that is distributed across multiple hospitals or medical devices, which are really concerned about their data privacy. The goal is to develop an ML model that can generalize well to a new patient data from a different hospital and can be used to make predictions or assist in clinical decision making. In this work, we train a convolutional neural network (CNN) model under decentralized FL setting on a medical dataset of OCT images for classification of patients with different eye diseases. We then compare its performance with the same CNN under the traditional centralized ML scheme where the entire dataset is available at one place, e.g., a hospital or a medical institution. We show that the decentralized FL model performs much worse (Accuracy=72.93%) than the centralized ML model (Accuracy=96.59%). However, with the use of local batch normalization layers in a global setting, the FL model can attain an improved accuracy of 86.16%. Hence, our results show promise towards achieving the overall goal of preserving privacy along with accurate prediction of patients’ medical conditions from their hospital data for providing solutions to their clinical problems.