Xsurv

Xsurv implements gradient boosting decision tree methods (Extreme Gradient Boosting, XGBoost, and Light Gradient Boosting, LightGBM) to perform survival analysis on censored clinical and high-dimensional molecular data for biomarker discovery and prognostic modeling.

Key Features:

• Advanced Algorithms: Employs gradient boosting decision tree frameworks—XGBoost and LightGBM—for modeling survival outcomes with improvements in training efficiency, scalability, and predictive accuracy.
• Cox Partial Likelihood Loss: Uses the partial likelihood function of the Cox proportional hazards model as a loss function to handle censored survival data.
• Smoothed C-index Loss Function: Incorporates a smoothed concordance index (C-index) as an alternative loss function for optimizing and evaluating predictive accuracy in survival models.
• Additional Modeling Options: Supports gradient boosting via the gbm package and random forests as alternative modeling approaches.
• Benchmarking and Performance: Benchmarked against stepwise Cox regression models and the original gbm gradient boosting implementation using comprehensive simulations, demonstrating improved handling of censored survival data.

Scientific Applications:

• Biomarker Discovery: Enables identification of prognostic candidate biomarkers from high-dimensional molecular data.
• Cancer Translational and Clinical Research: Facilitates prognostic modeling and outcome prediction in cancer studies using censored clinical data.
• Real-world Data Analysis: Has been applied to datasets such as melanoma methylation data to analyze complex biological signals related to survival.

Methodology:

Uses XGBoost and LightGBM GBDT frameworks; implements Cox partial likelihood and smoothed C-index as loss functions; supports gbm and random forests; benchmarked against stepwise Cox regression and gbm via comprehensive simulations and applied to melanoma methylation data.