Evaluating Random Forests for Survival Analysis using Prediction Error Curves.

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Prediction error curves are crucial for evaluating predictive performance in survival analysis.
• Assessing diverse prediction models, including traditional and machine learning approaches, requires robust computational tools.

Purpose of the Study:

• To survey the R package pec for efficient prediction error curve computation.
• To demonstrate extending pec's functionality to new prediction models, exemplified by random forests.
• To compare random forests with Cox regression using real-world datasets.

Main Methods:

• Utilizes inverse probability of censoring weights for right-censored data.
• Implements cross-validation techniques to address apparent error.
• Extends pec to support random forest models using the randomSurvivalForest and party R packages.

Main Results:

• The pec package facilitates efficient and flexible prediction error curve computation.
• Demonstrated successful integration of random forest models into the pec framework.
• Comparative analysis showed random forests as a viable alternative to Cox regression in specific settings.

Conclusions:

• The pec package is a valuable tool for comprehensive prediction error assessment in survival analysis.
• Extensibility of pec allows for the evaluation of novel and complex prediction models.
• Random forests offer a promising non-parametric approach for survival data analysis.