Feature selection methods for Cox proportional hazards model. Comparative study for financial and medical survival data

This study compares Cox proportional hazards models across medical and financial datasets built using various feature selection techniques. In this analysis 8 feature selection techniques (3 variants of forward selection, 2 variants of a selection based on principal component analysis, selection based on random survival forest, best subset selection and a selection based on a LASSO regularization) were tested across 22 multidimensional datasets (2 financial and 22 medical). The resulting Cox models were compared based on a concordance index. The main hypothesis of this study stating that the LASSO regularization or the selection based on random survival forest method (generating good models for medical data) would yield similar performance on financial data, was hereby disproved. The forward Schwarz and best subset selection gave the best results for financial data, while LASSO and random survival forest proved to be the most efficient in medical setting, for each considered model size.