Additive survival least square support vector machines and feature selection on health data in Indonesia

Abstract

Survival analysis is widely applied in many areas such as medicine, public health, engineering, economics, demography, and others. The initial approach has been employed for survival analysis is parametric model. Next, the semi parametric approach so-called Cox Proportional Hazard model (Cox-PHM) was developed. The parameters estimation in Cox PHM use partial likelihood function. The drawback of the Cox PHM is that it requires proportional condition in its hazard function between categories. In addition, it assumes linearity on its covariate pattern. This work employs nonparametric model so-called Additive Survival Least Square Support Vector Machines (A-SURLSSVM). The Cox PHM is used as a benchmark. The first data used in this study are generated from simulation. The second data are three health datasets in Indonesia. The performance of the proposed approach is compared with the benchmark based on the Concordance index (C-index) criterion. The higher C-index indicates better performance. In this study, application on three health datasets produce empirical results that conclude A-SURLSSVM perform better than Cox PHM for both with and without feature selection. In addition, the results of simulation study using 100 replications inform that feature selection can increase the C-index significantly. Moreover, the interaction between covariates yields the main confounder variable (the greatest probability to persist in the model) and the sub-main confounder (the most frequently excluded covariates from the model).