Cross-Silo Federated Learning for Predicting Successful Mechanical Ventilation Weaning Across 5 Intensive Care Unit Databases: Retrospective Database Analysis.

backgroundThe prediction of weaning from mechanical ventilation (MV) can support clinical decision-making and help reduce the risk of weaning failure in intensive care units (ICUs). Cross-silo federated learning (FL) offers a promising approach to developing robust predictive models across multiple institutions without requiring the sharing of patient-level data.

objectiveThis study aimed to evaluate the feasibility and efficacy of FL for predicting successful weaning from MV across 5 diverse ICU databases and to compare its performance with local learning (LL) and centralized learning (CL) approaches that differ in their data-sharing requirements.

methodsWe conducted a retrospective analysis using 5 ICU databases, namely the eICU Collaborative Research Database (eICU-CRD), Medical Information Mart for Intensive Care IV (MIMIC-IV), Universitätsklinikum Augsburg (UKA), High-Resolution ICU Dataset (HiRID), and Amsterdam University Medical Centers (AUMC), transforming clinical variables into the Observational Medical Outcomes Partnership (OMOP) Common Data Model. We defined successful weaning as a sustained reduction in positive end-expiratory pressure. We compared 3 learning approaches, FL, LL, and CL, using extreme gradient boosting (XGBoost). Performance was evaluated using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), precision, recall, and F

resultsA total of 24,521 patients were included across 5 databases. The CL model achieved an AUROC of 0.81, AUPRC of 0.57, and F

conclusionsOur findings highlight performance differences between learning approaches for MV weaning prediction. LL models achieved the highest performance within their respective institutions (AUROC=0.68-0.84), CL achieved the highest performance on pooled test data (AUROC=0.81), and FL showed lower but reasonable performance while avoiding direct sharing of patient-level data (AUROC=0.74). The choice between approaches depends on institutional data-sharing constraints, local dataset characteristics, and acceptable performance thresholds. As privacy was not formally measured, these findings should be read as a performance comparison rather than a privacy-performance trade-off.