A random forest approach for competing risks based on pseudo-values.

Area of Science:

• Machine Learning
• Biostatistics
• Survival Analysis

Background:

• Random forests are powerful supervised learning tools for prediction.
• Survival analysis with competing risks presents challenges due to censored data.
• Accurate event risk prediction is crucial in many medical and scientific fields.

Purpose of the Study:

• To extend the random forest method for event risk prediction in survival analysis with competing risks.
• To address the issue of right-censored data in prediction models.
• To introduce a novel pseudo-random forest approach using jackknife pseudo-values.

Main Methods:

• Developed a pseudo-random forest method by replacing censored event statuses with jackknife pseudo-values.
• Utilized node variance as a split criterion for regression trees due to continuous pseudo-responses.
• Compared the pseudo split criterion with the Gini split criterion in simulation studies.
• Evaluated predictive performance against Cox regression and random survival forest.

Main Results:

• The pseudo-random forest method demonstrated effective handling of censored data in competing risks scenarios.
• Simulation studies indicated competitive or superior predictive performance compared to existing methods.
• The method was successfully illustrated on two real-world datasets.

Conclusions:

• The proposed pseudo-random forest method offers a robust extension for risk prediction in survival analysis with competing risks.
• This approach provides a valuable tool for analyzing complex survival data, especially when dealing with censored observations.
• Further application and validation in diverse datasets are warranted.