JMIR Publications

ABSTRACT

Background:

Weaning from mechanical ventilation (MV) is a critical step in the care of intensive care unit (ICU) patients. Accurate prediction of successful weaning can support clinical decision-making, reduce complications, and optimize resource utilization. However, developing robust predictive models across multiple institutions is challenged by data privacy constraints. Cross-silo federated learning (FL) enables model training on decentralized data, preserving patient privacy while potentially improving generalizability across diverse clinical settings.

Objective:

To evaluate the feasibility and effectiveness of federated learning for predicting successful weaning from mechanical ventilation using data from multiple ICU databases, and to compare its performance against local and centralized learning approaches.

Methods:

This retrospective study analyzed data from five large ICU databases: 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). All data were harmonized into the Observational Medical Outcomes Partnership Common Data Model. Successful weaning was defined as a sustained reduction in positive end-expiratory pressure. We implemented and compared three learning strategies using XGBoost: federated learning, local learning at each site, and centralized learning on pooled data. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), precision, recall, and F1-score. All data use adhered to local ethical standards and IRB approvals.

Results:

A total of 24,521 patients were included across the five databases. The centralized learning approach yielded the highest overall performance (AUROC: 0.81, AUPRC: 0.57, F1-score: 0.56), closely followed by the federated model (AUROC: 0.80, AUPRC: 0.55, F1-score: 0.53). Local model performance varied substantially across sites (AUROC: 0.68–0.83, AUPRC: 0.52–0.71, F1-score: 0.51–0.68), reflecting data heterogeneity and differences in class balance. While centralized and federated models provided more consistent results, local models outperformed both on their respective datasets.

Conclusions:

Federated learning is a feasible and privacy-preserving approach for mechanical ventilation weaning prediction, offering competitive performance with only modest degradation compared to centralized learning. However, locally trained models showed the best performance on their own data, highlighting the benefits of data-specific tuning. The optimal approach depends on institutional data sharing policies, privacy requirements, and acceptable trade-offs between accuracy and privacy. Federated learning represents a viable middle ground when centralized data aggregation is not feasible.