Federated Learning for Early Detection of Neonatal Sepsis with Enhanced Data Interoperability
Abstract
Healthcare systems are increasingly dependent on data-driven approaches in patient care. However, heterogeneity and fragmentation of healthcare data cause problems for interoperability and seamless data integration across various healthcare sources. This study uses a synthetic dataset of late-onset neonatal sepsis and a public dataset of early-onset neonatal sepsis to discuss the potential influence of federated learning on healthcare interoperability. Federated learning provides a decentralized approach for training machine learning models across multiple institutions while maintaining the privacy and security of the data. This study demonstrates that integrating disparate healthcare sources using federated learning resulted in an accuracy of 65% for the late-onset neonatal sepsis dataset and 92% for the early-onset neonatal sepsis dataset, compared to 67% and 94% achieved without federated learning. Our findings will, therefore, be instrumental in deepening the understanding of the potential of federated learning to solve the interoperability problems of healthcare systems, paving the way for more efficient and effective data-driven solutions for healthcare.