Learning from vertically distributed data across multiple sites: An efficient privacy-preserving algorithm for Cox...

Area of Science:

• Distributed computing
• Biostatistics
• Machine learning

Background:

• Federated learning enables collaborative model training without direct data sharing.
• Distributed data presents challenges for traditional statistical modeling, particularly for survival analysis.
• Privacy-preserving methods are crucial for handling sensitive health data.

Purpose of the Study:

• To develop a lossless, distributed algorithm for the regularized Cox proportional hazards model with variable selection.
• To support federated learning for vertically distributed data in survival analysis.
• To enable accurate time-to-event data modeling without compromising patient privacy.

Main Methods:

• A novel distributed algorithm based on cyclical coordinate descent is proposed.
• Intermediary statistics are computed locally and exchanged for model updates, preserving data privacy.
• The algorithm was evaluated using simulations and real-world Alzheimer's dementia risk prediction data (ROSMAP).

Main Results:

• The distributed algorithm achieves privacy-preserving variable selection for time-to-event data with no accuracy loss compared to centralized methods.
• Simulations confirm high efficiency in analyzing high-dimensional datasets.
• Real-world analysis showed improved Alzheimer's dementia risk prediction accuracy and computational efficiency over existing privacy-preserving models.

Conclusions:

• The developed algorithm is lossless, privacy-preserving, and efficient for fitting regularized Cox models with distributed data.
• It offers a practical solution for distributed time-to-event data modeling.
• This approach facilitates collaborative research while maintaining data security and analytical integrity.