Learning from electronic health records across multiple sites: A communication-efficient and privacy-preserving...

Area of Science:

• Biostatistics
• Distributed Computing
• Health Informatics

Background:

• Logistic regression is crucial for analyzing binary outcomes in clinical research.
• Distributed algorithms are needed to analyze multi-site health data while preserving patient privacy.
• Current methods often require data pooling or iterative communication, posing privacy and efficiency challenges.

Purpose of the Study:

• To introduce a novel one-shot, privacy-preserving distributed algorithm for logistic regression (ODAL).
• To enable accurate logistic regression analysis across multiple clinical sites without transferring patient-level data.
• To evaluate the performance and efficiency of ODAL compared to traditional methods.

Main Methods:

• Proposed ODAL algorithm incorporating first-order (ODAL1) and second-order (ODAL2) gradients.
• Evaluated ODAL using extensive simulation studies.
• Applied ODAL to a real-world dataset from the University of Pennsylvania Health System to study medication effects on fetal loss.

Main Results:

• ODAL1 demonstrated a relative estimation bias <3% and a standard error ratio <1.25 compared to pooled data in simulations.
• ODAL2 achieved higher accuracy with relative bias <0.1% and standard error ratio <1.05.
• In real data analysis, ODAL2 showed <10% relative bias for 99% of medications studied.

Conclusions:

• ODAL is a privacy-preserving and communication-efficient method for distributed logistic regression.
• The algorithm offers small bias and high statistical efficiency, making it suitable for multi-site clinical data analysis.
• ODAL provides a viable alternative to data pooling for sensitive health information.