Performance analysis of the Survival-SVM classifier applied to gene-expression databases

Abstract

The analysis of epigenetic information for the diagnosis and prognosis of patients has been gaining relevance in recent years due to the technological progress that entails a decrease in information extraction and processing costs. One of the tasks most commonly carried out in this area is obtaining models that allow using patient epigenetic information to make inferences about survival analysis. As a result, optimizing these models turns into a problem of great interest today. In this article, the evaluation of different metrics and execution times for the Survival Support Vector Machines model is carried out through survival analysis applied to gene expression databases. Different experiments were performed varying the number of genes used for training to measure the correlation between model performance and data growth. The results showed that linear and polynomial kernels offer a better balance between execution time and model predictive power when the number of genes to be evaluated is less than 2000, while the cosine and RBF kernels are better candidates otherwise.