A Modified Random Survival Forests Algorithm for High Dimensional Predictors and Self-Reported Outcomes.

Area of Science:

• Biostatistics
• Epidemiology
• Bioinformatics

Background:

• High-dimensional datasets in epidemiology often feature time-to-event outcomes with measurement error.
• Accurate variable selection is crucial for identifying risk factors in large-scale studies like the Women's Health Initiative.
• Imperfect outcome data is also common in electronic medical records.

Purpose of the Study:

• To develop and evaluate an ensemble tree-based algorithm for variable selection with error-prone time-to-event outcomes.
• To address challenges in analyzing observational data where outcomes are not precisely recorded.
• To facilitate the discovery of significant variables in complex datasets.

Main Methods:

• An ensemble tree-based algorithm was developed for variable selection.
• The algorithm handles high-dimensional data with time-to-event outcomes observed with error.
• Performance was assessed using simulation studies with continuous and categorical covariates.

Main Results:

• The proposed algorithm demonstrates effectiveness in variable selection for datasets with imperfect time-to-event outcomes.
• Simulation studies confirmed the algorithm's performance across different covariate types.
• The approach successfully identified single nucleotide polymorphisms associated with incident Type II diabetes in the Women's Health Initiative cohort.

Conclusions:

• The developed ensemble tree-based algorithm provides a robust method for variable selection in the presence of outcome error.
• This approach is applicable to various settings, including large epidemiologic studies and electronic medical record data.
• An R package, 'icRSF', is available for implementing these novel methods.