Federated Mixed Effects Logistic Regression Based on One-Time Shared Summary Statistics.

Area of Science:

• Biostatistics
• Medical Informatics
• Data Privacy

Background:

• Accessing individual-level patient data for medical research is challenging due to privacy concerns.
• Estimating complex statistical models like mixed effects binary logistic regression across multiple data sources (e.g., hospitals) is difficult while maintaining data privacy.

Purpose of the Study:

• To develop a novel strategy for estimating mixed effects binary logistic regression models that preserves individual patient data privacy.
• To enable collaborative medical research by allowing multiple data providers to contribute to a global model without sharing raw data.

Main Methods:

• A novel federated learning approach is proposed, requiring data providers to share summary statistics only once.
• Pseudo-data generation is employed, where summary statistics mimic the actual data for model estimation.
• The method accommodates multiple predictors, including continuous and categorical variables.

Main Results:

• The proposed strategy accurately estimates the mixed effects binary logistic regression model, performing comparably to methods using pooled individual observations.
• Simulations demonstrate the effectiveness of the approach in various scenarios.
• An illustrative example using real-world medical data is provided.

Conclusions:

• This method offers a communication-efficient and secure alternative to traditional federated learning for statistical modeling in medical research.
• It eliminates the need for extensive infrastructure and addresses security concerns associated with data sharing.
• The approach effectively accounts for heterogeneity across data providers while protecting patient privacy.