A gradient boosting decision tree based estimation method for the mixture cure model.

Area of Science:

• Biostatistics
• Survival Analysis
• Machine Learning in Medicine

Background:

• Cure models are essential for censored survival data with a cured fraction.
• Existing semiparametric methods require restrictive parametric assumptions.
• Nonparametric methods are limited to single covariates.

Purpose of the Study:

• To propose a novel gradient boosting decision tree (GBDT) based method for estimating mixture cure models.
• To overcome limitations of existing semiparametric and nonparametric approaches.
• To provide more accurate estimates for cure probability and relative risk.

Main Methods:

• Utilizing a gradient boosting decision tree framework for cure model estimation.
• Developing a method that accommodates complex covariate effects without a priori parametric assumptions.
• Leveraging GBDT's ability to handle high-dimensional data.

Main Results:

• The proposed GBDT method yields more accurate estimates of cure probability and relative risk compared to existing methods.
• Simulation studies with large samples show small mean square errors for cure probability, relative risk score, and survival function estimates.
• The method demonstrates potential for analyzing high-dimensional covariates in survival data.

Conclusions:

• The GBDT-based cure model offers a flexible and accurate alternative to traditional methods.
• This approach enhances survival data analysis, particularly in scenarios with complex covariate effects and high dimensionality.
• The method shows promise for applications like cancer survival studies, as illustrated with colon cancer data.