Related Experiment Videos
Cross-Silo Federated Learning for Predicting Successful Mechanical Ventilation Weaning Across 5 Intensive Care Unit
Seyedmostafa Sheikhalishahi1, Mathias Kaspar1, Johanna Schwinn1
1Digital Medicine, University Hospital Augsburg, Augsburg, Germany.
JMIR Medical Informatics
|August 14, 2026
Summary
Federated learning (FL) offers a privacy-preserving method for predicting mechanical ventilation (MV) weaning. While centralized learning (CL) performed best overall, FL showed reasonable results without sharing patient data.
Area of Science:
- Critical Care Medicine
- Machine Learning in Healthcare
- Data Privacy in AI
Background:
- Predicting mechanical ventilation (MV) weaning is crucial for clinical decisions in intensive care units (ICUs).
- Federated learning (FL) enables developing predictive models across institutions without sharing patient data, addressing privacy concerns.
Purpose of the Study:
- To assess the feasibility and effectiveness of FL for predicting successful MV weaning.
- To compare FL performance against local learning (LL) and centralized learning (CL) models.
Main Methods:
- Retrospective analysis of 5 diverse ICU databases (eICU-CRD, MIMIC-IV, UKA, HiRID, AUMC).
- Clinical variables standardized to OMOP Common Data Model.
- Comparison of FL, LL, and CL using XGBoost, evaluated by AUROC, AUPRC, precision, recall, and F1-score.
Main Results:
- Centralized learning (CL) achieved the highest performance on pooled data (AUROC=0.81).
- Federated learning (FL) demonstrated reasonable performance (AUROC=0.74) while preserving data privacy.
- Local learning (LL) showed variable performance across databases (AUROC=0.68-0.84).
Conclusions:
- Performance varies significantly between LL, CL, and FL approaches for MV weaning prediction.
- The optimal approach depends on institutional data-sharing policies, local data characteristics, and performance requirements.
- FL provides a viable alternative for multi-institutional model development when direct data sharing is restricted.