Robust Inference for Federated Meta-Learning.

Area of Science:

• Data Science
• Statistical Inference
• Machine Learning

Background:

• Synthesizing multi-source data is crucial for generalizable knowledge but faces challenges due to data heterogeneity and sharing restrictions.
• Federated meta-learning offers a solution by enabling collaborative model training across multiple sites without centralizing data.

Purpose of the Study:

• To develop a robust inference framework for federated meta-learning that facilitates statistical inference for the prevailing model across diverse data sources.
• To address the challenges of site selection uncertainty and data heterogeneity in federated learning settings.

Main Methods:

• A novel sampling method is proposed to manage the additional variation introduced by data-adaptive site selection.
• A confidence interval is developed that is valid without requiring error-free site selection and does not necessitate sharing of individual-level data.
• The robust inference for federated meta-learning (RIFL) methodology is demonstrated across various inference problems, including parametric model aggregation, high-dimensional prediction, and average treatment effect estimation.

Main Results:

• The RIFL methodology provides valid statistical inference for the prevailing model in federated meta-learning settings.
• The proposed confidence interval accounts for selection uncertainty without compromising data privacy.
• RIFL was successfully applied to federated learning of COVID-19 mortality risk using real-world EHR data from 15 healthcare centers.

Conclusions:

• RIFL offers a broadly applicable and robust framework for federated meta-learning, enhancing knowledge generalizability from multi-source data.
• The methodology effectively addresses data heterogeneity and sharing constraints, enabling reliable statistical inference.
• The application to COVID-19 mortality risk demonstrates the practical utility of RIFL in real-world healthcare scenarios.