Comparison of Random Survival Forest and Cox Model for Prediction Performance: A Case Study

Survival analysis is currently one of the fastest growing statistical tools in academia. In survival analysis there is a robust regression model theory that can be used to model data with incomplete observations called censoring, most of these models are parametric, and there is also the Cox proportional hazards model. Machine Learning in conjunction with Random Forest in Survival Analysis (RSF) are an increasing alternative for use in prediction. Four different configurations of coefficients were adjusted in the RSF, starting from a saturated model with presence of interaction to a parsimonious model based on the criteria of the Machine Learning area to choose variables. The models were compared against the Cox model using the C-index and Brier Score Index (IBS) criteria. The best model adjusted for prediction was the complete model with all covariates under Random Survival Forest modeling.