Enhancing healthcare data privacy and interoperability with federated learning

Abstract

This article explores the application of federated learning (FL) with the Fast Healthcare Interoperability Resources (FHIR) protocol to address the underutilization of the huge volumes of healthcare data generated by the digital health revolution, especially those from wearable sensors, due to privacy concerns and interoperability challenges. Despite advances in electronic medical records, mobile health applications, and wearable sensors, current digital health cannot fully exploit these data due to the lack of data analysis and exchange between heterogeneous systems. To address this gap, we present a novel converged platform combining FL and FHIR, which enables collaborative model training that preserves the privacy of wearable sensor data while promoting data standardization and interoperability. Unlike traditional centralized learning (CL) solutions that require data centralization, our platform uses local model learning, which naturally improves data privacy. Our empirical evaluation demonstrates that federated learning models perform as well as, or even numerically better than, centralized learning models in terms of classification accuracy, while also performing equally well in regression, as indicated by metrics such as accuracy, area under the curve (AUC), recall, and precision, among others, for classification, and mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE) for regression. In addition, we developed an intuitive AutoML-powered web application that is FL and CL compatible to illustrate the feasibility of our platform for predictive modeling of physical activity and energy expenditure, while complying with FHIR data reporting standards. These results highlight the immense potential of our FHIR-integrated federated learning platform as a practical framework for future interoperable and privacy-preserving digital health ecosystems to optimize the use of connected health data.