WebDISCO: a web service for distributed cox model learning without patient-level data sharing.

Area of Science:

• Biostatistics
• Computational Biology
• Health Informatics

Background:

• The Cox proportional hazards model is essential for survival data analysis but demands substantial data for statistical power.
• Sharing data across institutions is a potential strategy to increase power in survival analyses.
• Existing methods often require centralizing sensitive patient data, posing privacy concerns.

Purpose of the Study:

• To develop and validate a federated learning approach for Cox proportional hazards models.
• To enable distributed survival analysis without direct sharing of patient-level data.
• To demonstrate the technical feasibility of a web service for federated survival analysis.

Main Methods:

• Development of WebDISCO, a web service for distributed Cox model learning.
• Implementation of a federated algorithm where patient data is processed locally.
• Exchange of only intermediate statistics to construct a global Cox model.

Main Results:

• Mathematical derivation confirms the distributed algorithm yields results identical to centralized Cox models.
• Experimental evaluation at multiple institutions showed near-identical model coefficients between distributed and centralized approaches.
• The implementation demonstrated the ability to achieve consistent results with centralized methods.

Conclusions:

• WebDISCO serves as a proof-of-concept for federated survival analysis, demonstrating technical feasibility.
• The approach allows for distributed Cox model learning without compromising patient-level data privacy.
• This method provides a foundation for overcoming data limitations in survival analysis through federated learning.