BoXHED2.0: Scalable Boosting of Dynamic Survival Analysis

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Modern survival analysis frequently incorporates time-dependent covariates.
• Existing methods may not adequately address complex survival settings beyond simple right-censoring.

Purpose of the Study:

• Introduce BoXHED2.0, a novel Python package for advanced survival analysis.
• Provide a flexible and efficient tool for handling time-dependent covariates in various survival data scenarios.

Main Methods:

• BoXHED2.0 employs a tree-boosted, fully nonparametric hazard estimation approach.
• The core implementation in C++ allows for significant computational speedups.
• Supports parallel processing via GPUs and multicore CPUs for scalability.

Main Results:

• BoXHED2.0 is applicable to general survival settings, including recurring events and competing risks.
• Achieves computational performance comparable to parametric boosted survival models.
• Demonstrates scalability for large datasets and complex analyses.

Conclusions:

• BoXHED2.0 provides a powerful, nonparametric solution for survival analysis with time-dependent covariates.
• Its efficiency and flexibility make it suitable for a wide range of modern biostatistical and machine learning applications.
• The package is readily available for use and further development.