ODACH: a one-shot distributed algorithm for Cox model with heterogeneous multi-center data.

Area of Science:

• Biostatistics
• Health Informatics
• Epidemiology

Background:

• Multi-center studies generate valuable time-to-event data.
• Analyzing such data often requires sharing sensitive patient-level information.
• Existing methods like meta-analysis can be biased, especially with rare events.

Purpose of the Study:

• To develop a privacy-preserving distributed algorithm for analyzing heterogeneous multi-center time-to-event data.
• To circumvent the need for sharing patient-level data across research sites.
• To enable accurate Cox proportional-hazards modeling in distributed settings.

Main Methods:

• Developed the One-shot Distributed Algorithm for Cox proportional-hazards model (ODACH).
• Implemented a surrogate likelihood function to approximate the Cox log-partial likelihood.
• Stratified the model by site, using lead site data and aggregated site information.
• Allowed for site-specific baseline hazard functions and covariate distributions.

Main Results:

• ODACH provided estimates comparable to the pooled estimator, which uses direct patient-level data.
• The algorithm demonstrated reduced bias compared to meta-analysis, particularly for rare events.
• Simulation studies and a real-world opioid use disorder study validated ODACH's performance.

Conclusions:

• ODACH is an effective privacy-preserving method for analyzing multi-center time-to-event data.
• The algorithm is communication-efficient, suitable for distributed research networks.
• ODACH offers a robust alternative to traditional methods, enhancing data security and analytical accuracy.