Random Survival Forests With Competing Events: A Subdistribution-Based Imputation Approach.

Area of Science:

• * Biostatistics
• * Machine Learning
• * Survival Analysis

Background:

• * Random survival forests (RSF) are valuable for complex time-to-event data.
• * Competing events, common in clinical settings, can bias traditional RSF analyses if treated as censoring.
• * Fine & Gray's subdistribution hazard model offers an alternative for competing risks but is not directly integrated with RSF.

Purpose of the Study:

• * To integrate competing risk concepts into Random Survival Forests (RSF).
• * To develop and evaluate imputation strategies for handling competing events within an RSF framework.
• * To improve the estimation of the cumulative incidence function (CIF) in the presence of competing risks.

Main Methods:

• * Developed imputation strategies adapting discrete-time subdistribution hazard model weights.
• * Integrated these imputation methods into the Random Survival Forest algorithm.
• * Conducted simulations to assess the performance of the proposed approach.
• * Applied the method to an epidemiological dataset on chronic kidney disease.

Main Results:

• * Simulations demonstrated accurate CIF estimation when imputation is performed globally.
• * The proposed RSF approach effectively handles competing events, avoiding bias from treating them as censoring.
• * The method showed strong performance even with low event rates or high censoring.
• * Analysis of the chronic kidney disease dataset yielded plausible predictor-response relationships and CIF estimates.

Conclusions:

• * Integrating competing risk modeling into RSF provides a robust approach for time-to-event analysis.
• * The developed imputation strategies are effective in improving CIF estimation accuracy.
• * This enhanced RSF method is crucial for clinical and epidemiological research where competing events are prevalent.
• * Neglecting competing events can lead to significant biases in survival analyses.