Gradient boosting-based discrete failure time model for selecting time-varying effects and interactions

Area of Science:

• Biostatistics
• Survival Analysis
• Machine Learning

Background:

• Analyzing National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) Program data is crucial for cancer management.
• Challenges exist in variable selection, differentiating time-varying effects, and selecting interaction terms under hierarchy restrictions in survival models.

Purpose of the Study:

• To develop a novel coordinate ascent-based gradient boosting procedure for discrete failure time models.
• To address limitations in variable selection, time-varying effect differentiation, and interaction term selection in high-dimensional survival data.

Main Methods:

• A coordinate ascent-based gradient boosting procedure was developed for discrete failure time models.
• The method incorporates variable selection, distinguishes time-varying and time-independent effects, and handles interaction terms under hierarchy restrictions.
• It provides well-defined degrees of freedom for information-criteria-based stopping rules.

Main Results:

• The proposed method achieves effective variable selection in high-dimensional settings.
• It successfully differentiates between time-varying and time-independent variables and selects important interaction terms.
• Simulation studies demonstrated good selection performance, and application to SEER melanoma data identified key risk factors and interactions.

Conclusions:

• The developed gradient boosting procedure offers a robust solution for complex survival data analysis.
• It enhances the ability to identify prognostic factors and understand temporal effects in cancer research.
• This method provides improved tools for guiding cancer management strategies using SEER data.