L₁ splitting rules in survival forests

Area of Science:

• Machine Learning
• Biostatistics
• Survival Analysis

Background:

• The log-rank test is a standard splitting criterion in survival trees and forests.
• Log-rank test can exhibit reduced statistical power when survival functions cross or hazard functions differ.
• This limitation can impact the accuracy of survival prediction models.

Purpose of the Study:

• To evaluate an alternative splitting rule for survival forests.
• To compare the performance of forests using integrated absolute difference versus the log-rank test.
• To assess the effectiveness of the new rule in diverse data scenarios.

Main Methods:

• Investigated the integrated absolute difference between child nodes' survival functions as a splitting rule.
• Conducted simulation studies to compare the proposed rule against the log-rank test.
• Applied both splitting rules to real-world datasets for validation.

Main Results:

• Survival forests utilizing the integrated absolute difference rule generally yield superior results.
• The proposed rule demonstrates improved performance compared to the log-rank splitting rule in various settings.
• Effectiveness is particularly noted in scenarios where survival functions intersect.

Conclusions:

• The integrated absolute difference is a robust and effective splitting rule for survival forests.
• This alternative offers advantages over the log-rank test, especially in complex survival data.
• The findings suggest a valuable enhancement for survival tree and forest algorithms.