Abstract
OBJECTIVE: The objective of this retrospective study is to compare three learning approaches for blood transfusion (BT) prediction after intensive care unit (ICU) admission: local learning (LL), federated learning (FL), and centralized learning (CL) across five ICU databases (eICU Collaborative Research Database, Medical Information Mart for Intensive Care IV, High-Resolution Intensive Care Unit Dataset, Amsterdam University Medical Center Database, University Hospital of Augsburg). METHODS: As machine learning model we used XGBoost and included 15 clinical variables. The prediction task consists of a 3-h observation window, followed by a 2-h prediction window. We evaluated the models using internal and external validation with area under the receiver-operator curve, area under the precision-recall curve, PPV, Brier score and F1 score. RESULTS: CL consistently outperformed FL and LL in both internal validation (AUPRC range: 0.73–0.95 (CL) vs 0.63–0.96 (FL) and 0.69–0.96 (LL)) and external validation (AUPRC range: 0.61–0.89 (CL) vs 0.45–0.91 (FL) and 0.37–0.90 (LL)). FL showed variable performance across datasets. CONCLUSIONS: The complexity of the multivariable clinical prediction of BTs may create substantial challenges for FL effectiveness, particularly under high data heterogeneity conditions that are common in healthcare.