Federated horizontally partitioned principal component analysis for biomedical applications

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• Federated learning (FL) enables privacy-preserving machine learning in medicine by keeping sensitive patient data decentralized.
• FL introduces challenges like batch effects and population heterogeneity in medical data analysis.
• Principal Component Analysis (PCA) is a key tool for machine learning and data visualization.

Purpose of the Study:

• To investigate challenges in adapting classical Principal Component Analysis (PCA) for federated learning scenarios.
• To implement and evaluate federated PCA algorithms for high-dimensional biological data.
• To assess the accuracy and utility of federated PCA for downstream analyses in decentralized medical datasets.

Main Methods:

• Implementation of various federated PCA algorithms.
• Evaluation using high-dimensional biological data with realistic multi-site distributions.
• Assessment of accuracy against centralized PCA solutions and impact on downstream analyses.

Main Results:

• Federated subspace iteration converges to the centralized solution, even with challenging data distributions.
• Approximate federated PCA methods introduce errors, but accuracy improves with larger sample sizes.
• Approximate methods are suitable for basic data visualization but sensitive to outliers and batch effects.

Conclusions:

• Federated subspace iteration is a robust method for decentralized PCA in medical research.
• Careful selection of PCA algorithms and parameters is crucial to minimize communication overhead.
• Federated PCA methods can be effectively applied to preserve privacy in large-scale biological data analysis.