Federated double machine learning for high-dimensional semiparametric models.

Area of Science:

• Machine Learning
• Statistical Modeling
• Biostatistics

Background:

• Federated learning trains global models with decentralized data but struggles with high-dimensional semiparametric models and nuisance parameters.
• Multicenter studies present unique challenges due to data heterogeneity and the need for privacy-preserving analysis.

Purpose of the Study:

• To propose a novel federated double machine learning framework for semiparametric models in multicenter studies.
• To address challenges posed by high-dimensional nuisance parameters in federated settings.
• To develop a robust federated estimator combining local and aggregated data.

Main Methods:

• Leveraging double machine learning (DML) for robust parameter estimation.
• Extending the surrogate efficient score method within a Neyman-orthogonal framework.
• Applying density ratio tilting for federated estimator construction, integrating individual data with summary statistics.

Main Results:

• The proposed methodology effectively mitigates regularization bias and overfitting in high-dimensional nuisance parameter estimation.
• The estimator's limiting distribution is established under minimal assumptions.
• Validated through extensive simulations and real-world application to Alzheimer's Disease Neuroimaging Initiative data.

Conclusions:

• The federated double machine learning framework offers a powerful solution for analyzing complex data in multicenter studies.
• This approach enhances the accuracy and reliability of federated learning for semiparametric models.
• Demonstrates significant potential for applications in various fields requiring privacy-preserving, high-dimensional data analysis.