CEDAR: communication efficient distributed analysis for regressions.

Area of Science:

• Health Informatics
• Biostatistics
• Machine Learning

Background:

• Electronic health records (EHRs) hold significant potential for precision medicine.
• Sharing patient-level data across institutions is restricted by regulations and policies.
• Distributed learning over multiple EHR databases without data sharing is a growing area of interest.

Purpose of the Study:

• To propose a communication-efficient distributed learning method for EHRs.
• To enable statistical inference without sharing raw patient-level data.
• To ensure differential privacy and reduce information leakage risks.

Main Methods:

• Framing the distributed learning problem as a missing data problem.
• Aggregating optimal estimates from external sites.
• Incorporating posterior samples from remote sites for improved efficiency.
• Theoretical investigation of asymptotic properties for statistical inference and differential privacy.

Main Results:

• The proposed method allows for proper statistical inference without raw data sharing.
• Demonstrated improved efficiency of parameter estimates through posterior sample incorporation.
• Achieved differential privacy, mitigating risks of information leakage.
• Evaluated performance via simulations and real-world data analyses.

Conclusions:

• The novel distributed learning approach effectively addresses EHR data sharing challenges.
• The method provides a secure and efficient framework for multi-institutional EHR analysis.
• This work advances precision medicine through privacy-preserving distributed statistical inference.