Addressing missing values and fragmented data to improve core and management processes in hospitals: a federated machine learning approach

Abstract Despite the increasing digitization of healthcare data, which offers opportunities to enhance core and management processes, hospitals frequently encounter challenges related to missing data and data fragmentation across institutions. In healthcare, missing data arise from both human and technological errors. Additionally, fragmented health data are often difficult to transfer due to their highly sensitive nature, which is subject to strict ethical and legal regulations. This study builds on the state-of-the-art Federated Machine Learning technique, Federated Averaging. We introduce two novel Federated Learning approaches: The Flexible and the Modified Federated Stochastic Variance Reduced Gradient (F-FSVRG and M-FSVRGS). These methods are designed to address missing and fragmented data with increased flexibility and robustness. We evaluate the performance of F-FSVRG and M-FSVRGS through a computational study using real-world multicenter COVID-19 diagnosis and Intensive Care Unit (ICU) data. M-FSVRGS generally achieves equal or higher accuracy than F-FSVRG, mostly independent of dataset size and structure, with statistically significant improvements observed for the ICU dataset when the proportion of missing values is high. Our findings demonstrate that M-FSVRGS effectively mitigates the challenges of fragmented and missing data, facilitating the integration of AI-driven approaches. Furthermore, this advancement supports sustainable improvements in healthcare processes and operations.