Linear Mixed Modeling of Federated Data When Only the Mean, Covariance, and Sample Size Are Available

Area of Science:

• Medical Statistics
• Health Informatics
• Biostatistics

Background:

• Patient data confidentiality is paramount in medical research.
• Statistical modeling with decentralized data presents privacy challenges.
• Federated learning offers a solution but requires iterative communication.

Purpose of the Study:

• To propose an alternative framework to federated learning for fitting linear mixed models.
• To enable accurate statistical analysis while preserving patient confidentiality.
• To develop a communication-efficient method for decentralized data analysis.

Main Methods:

• Developed a novel framework utilizing summary statistics (mean, covariance, sample size) from data providers.
• Leveraged the principle of statistical sufficiency within the likelihood framework.
• Applied the method to model COVID-19 PCR test cycle threshold using real patient data.

Main Results:

• Achieved estimates identical to those derived from individual-level data.
• Demonstrated the approach on 15,068 patient records from 70 clinics.
• The method requires only a single communication of summary statistics per data provider.

Conclusions:

• The proposed approach offers a simpler, more communication-efficient alternative to federated learning for linear mixed models.
• It ensures patient data confidentiality by only sharing aggregated summary statistics.
• The framework is generalizable and can be implemented across various statistical software.