Privacy-Preserving Federated Learning for Healthcare Data Classification

Abstract

This study proposes a federated learning model for patient health record classification without sharing of sensitive information among healthcare institutions. The anonymized data set, consisting of both demographic and clinical data and diagnostic markers, was shared among five simulated health care clients who independently trained local neural-network classification model for a fixed number of communication rounds. Instead of sending patient information, only the model parameters were sent locally to a central aggregation server, where the models of the participating clients were federated averaged to form a global model, which was then sent back to the participating clients. To find out whether federated learning could provide similar accuracy levels to conventional centralized machine learning while minimizing the direct exposure of sensitive healthcare information, the proposed federated approach was compared with the conventional centralized machine learning approach, and their comparative performance was evaluated in terms of accuracy, precision, recall, F1-score, and communication overhead. Results showed that the federated global model performed well in terms of classification accuracy, with its performance getting closer to the centralized baseline with more communication rounds. The total data sent across all rounds – which includes data sent from the client to the local site and from the local site to the federation – was significantly lower for the federated approach than it would be if the raw data were gathered in one place and sent to each client during training, while no patient data were ever sent to any client outside the local site during training. The study shows that federated learning with federated averaging is a feasible approach to meet a healthcare classification accuracy close to centralized training without compromising on the amount of sensitive patient data shared, and provides suggestions for healthcare data governance, deployment and future studies.