A Comparative Analysis of Federated and Centralized Learning for SpO2 Prediction in Five Critical Care Databases.

ABSTRACT

This study explores the potential of federated learning (FL) to develop a predictive model of hypoxemia in intensive care unit (ICU) patients. Centralized learning (CL) and local learning (LL) approaches have been limited by the localized nature of data, which restricts CL approaches to the available data due to data privacy regulations. A CL approach that combines data from different institutions, could offer superior performance compared to a single-institution approach. However, the use of this method raises ethical and regulatory concerns. In this context, FL presents a promising middle ground, enabling collaborative model training on geographically dispersed ICU data without compromising patient confidentiality. This study is the first to use all five public ICU databases combined. The findings demonstrate that FL achieved comparable or even slightly improved performance compared to local or centralized learning approaches.