Is Random Survival Forest an Alternative to Cox Proportional Model on Predicting Cardiovascular Disease?

Random survival forest (RSF), a non-parametric and non-linear approach for survival analysis, has been used in several risk models and presented to be superior to traditional Cox proportional model. Anyway, can RSF replace Cox proportional model on predicting cardiovascular disease? In this paper, we evaluate the performance of RSF by comparing it with Cox in terms of discrimination ability, ability to identify non-linear effects and ability to identify important predictors that can discriminate survival function. Two databases are studied, including heart failure population database and cardiac arrhythmias database. We take 1-year mortality after cardiac arrhythmias prediction as an example for comparison between Cox and RSF based model. The results show that RSF improved discrimination performance greatly than Cox with an out-of-bag C-statistics of 0.812 (while 0.736 for Cox based model). In addition, RSF can automatically identify non-linear effects of all variables but Cox cannot. However, RSF is inferior in identifying predictors with less ratio of population due to its insensitivity to noise. Therefore, RSF cannot replace Cox in current status and should be studied further.