Federated Learning for Predictive Analytics in Weaning from Mechanical Ventilation.

Journal: Studies in health technology and informatics
Published Date:

Abstract

Mechanical ventilation is crucial for critically ill patients in ICUs, requiring accurate weaning and extubations timing for optimal outcomes. Current prediction models struggle with generalizability across datasets like MIMIC-IV and eICU-CRD. We propose a federated learning approach using XGBoost with bagging aggregation to improve weaning predictions while ensuring patient data privacy, compliant with GDPR and HIPAA. Using the OMOP Common Data Model, our method integrates machine learning techniques across three ICU databases, encompassing over 33,000 patients. Our model achieved robust performance with 77% AUC and 73% AUPRC. Planned pilot studies in Germany will further refine and validate our approach. This study demonstrates the potential of federated learning to enhance critical care by providing personalized, data-driven insights for ventilation management.

Authors

  • Seyedmostafa Sheikhalishahi
    University of Trento, Trento, Italy.
  • Johanna Schwinn
    Digital Medicine, University Hospital of Augsburg, Augsburg, Germany.
  • Matthaeus Morhart
    Digital Medicine, University Hospital of Augsburg, Augsburg, Germany.
  • Mathias Kaspar
    University Hospital of Würzburg, Comprehensive Heart Failure Center.
  • Ludwig Christian Hinske
    Institute for Digital Medicine, University Hospital Augsburg, Augsburg, Germany.