FEDERATED LEARNING OF ROBUST INDIVIDUALIZED DECISION RULES WITH APPLICATION TO HETEROGENEOUS MULTIHOSPITAL SEPSIS...

Area of Science:

• Computational biology
• Health informatics
• Machine learning in healthcare

Background:

• Sepsis affects millions annually, necessitating personalized treatment strategies.
• Electronic health records from multiple hospitals offer rich data for sepsis management.
• Existing methods struggle with data heterogeneity across hospitals, limiting generalizability of decision rules.

Purpose of the Study:

• To develop individualized decision rules (IDRs) for sepsis management adaptable across diverse hospital settings.
• To address data heterogeneity and data sharing restrictions in multi-hospital electronic health record data.
• To enhance sepsis patient outcomes through robust, universally applicable decision-making tools.

Main Methods:

• Introduced a novel conditional maximin objective function for robust IDR learning.
• Developed a federated learning algorithm to handle distributional uncertainty from heterogeneous data.
• Trained and validated IDRs using electronic health records from multiple UPMC hospitals, focusing on data privacy.

Main Results:

• The proposed method significantly enhances survival rates, particularly for high-risk patients (10 percentage point increase).
• Overall survival rates improved by 2-3 percentage points when applied to unseen hospital populations.
• Demonstrated robustness of IDRs against hospital-level variations and distributional shifts.

Conclusions:

• Federated learning with a conditional maximin objective offers a robust framework for learning generalizable IDRs in sepsis.
• This approach effectively addresses data heterogeneity and privacy concerns in multi-institutional healthcare data.
• The developed IDRs have the potential to uniformly improve sepsis management and patient outcomes across healthcare systems.