Survival Prediction with Extreme Learning Machine, Supervised Principal Components and Regularized Cox Models in High-Dimensional Survival Data by Simulation.

Mortality risks of important diseases such as cancer can be estimated using gene profiles which are high-dimensional data obtained from gene expression sequences. However, it is impossible to analyze high-dimensional data with classical techniques due to multicollinearity, timeconsuming processing load, and difficulty interpreting the results. For this purpose, extreme learning machine methods, which can solve regression and classification problems, have become one of the most preferred machine learning methods regarding fast data analysis and ease of application. The goal of this study is to compare estimation performance of risk score and shortterm survival with survival extreme learning machine methods, L2-penalty Cox regression, and supervised principal components analysis in generated high-dimensional survival data. The survival models have been evaluated by Harrell's concordance index, integrated Brier score, F1 score, kappa coefficient, the area under the curve, the area under precision-recall, accuracy, and Matthew's correlation coefficient. Performances of risk score estimation and short-term survival prediction of the survival models for the censoring rates of 10%, 30%, 50% and 70% have been obtained in the range of 0.746-0.796, 0.739-0.798, 0.726-0.791, 0.708-0.784 for Harrell's concordance index; 0.773-0.824, 0.772-0.824, 0.754-0.818, 0.739-0.808 for F1 score and 0.816-0.867, 0.808-0.865, 0.788-0.863, 0.776-0.851 for area under curve. All results showed that survival extreme learning machine methods that allow analyzing high-dimensional survival data without the necessity of dimension reduction perform very competitive with the other popular classical methods used in the study. [ABSTRACT FROM AUTHOR]