Federated Learning for Predictive Analytics in Weaning from Mechanical Ventilation.

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Data Privacy in Research

Background:

• Mechanical ventilation is vital for intensive care unit (ICU) patients.
• Accurate timing for weaning and extubation is critical for patient outcomes.
• Existing prediction models lack generalizability across different ICU datasets.

Purpose of the Study:

• To develop a generalizable machine learning model for predicting mechanical ventilation weaning.
• To ensure patient data privacy using a federated learning approach.
• To improve the accuracy of extubation timing predictions.

Main Methods:

• Federated learning utilizing XGBoost with bagging aggregation.
• Integration of machine learning across three ICU databases using the OMOP Common Data Model.
• Analysis of data from over 33,000 patients, ensuring GDPR and HIPAA compliance.

Main Results:

• Achieved robust performance with an Area Under the Curve (AUC) of 77%.
• Obtained an Area Under the Precision-Recall Curve (AUPRC) of 73%.
• Demonstrated the feasibility of cross-institutional collaboration while preserving data privacy.

Conclusions:

• Federated learning offers a promising solution for developing generalizable predictive models in critical care.
• The proposed method enhances personalized, data-driven insights for mechanical ventilation management.
• Future pilot studies in Germany will further validate and refine the approach.