Regression trees and ensembles for cumulative incidence functions.

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Cumulative incidence functions are vital for understanding event risks in the presence of competing events.
• Traditional methods include parametric, nonparametric, and semi-parametric approaches.
• Machine learning extensions, like regression trees, are emerging for competing risks analysis.

Purpose of the Study:

• To propose a novel approach for estimating cumulative incidence curves using regression trees and ensemble methods.
• To develop and implement easily usable methods for competing risks data analysis.
• To apply these new methods to real-world clinical trial data.

Main Methods:

• Utilized regression trees and ensemble estimators for competing risks modeling.
• Employed augmented estimators of the Brier score risk for tree building and pruning.
• Leveraged existing R packages for practical implementation.

Main Results:

• The proposed regression tree and ensemble methods provide a new framework for estimating cumulative incidence.
• The methods are shown to be implementable using standard statistical software.
• Illustrative analysis using Radiation Therapy Oncology Group (trial 9410) data demonstrates the utility of the approach.

Conclusions:

• Novel machine learning-based methods for cumulative incidence estimation in competing risks settings have been developed.
• These methods offer a practical and implementable alternative to traditional statistical techniques.
• The approach shows promise for analyzing complex event data in clinical research.