A gradient boosting algorithm for survival analysis via direct optimization of concordance index.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Survival analysis models often rely on restrictive assumptions about hazard functions.
• The proportional hazards model assumes a multiplicative relationship between covariates and hazard.
• There is a need for flexible, assumption-free survival analysis methods.

Purpose of the Study:

• To propose a novel nonparametric survival analysis model.
• To develop a model that does not impose specific forms on hazard functions.
• To evaluate the performance of the proposed model against existing survival analysis methods.

Main Methods:

• Utilized an ensemble of regression trees for nonparametric hazard function estimation.
• Employed gradient boosting for model training, optimizing a smoothed concordance index approximation.
• Implemented the model in R as the Gradient Boosting Machine for Concordance Index (GBMCI) software package.

Main Results:

• The GBMCI model demonstrated consistent performance improvements over other survival models.
• Benchmarking was conducted using a large-scale breast cancer prognosis dataset.
• The model's effectiveness was validated across various covariate settings.

Conclusions:

• The proposed nonparametric survival analysis model (GBMCI) offers superior predictive performance.
• GBMCI provides a flexible alternative to traditional survival models with strong assumptions.
• The GBMCI R package is freely available, facilitating broader application in survival data analysis.