Comparative analysis of high dimensional survival data with supervised principal components, penalized cox regression and extreme learning machines methods

Objective: The goal of the study is to compare the performances of extreme learning machines-based survival, supervised principal components analysis, and L2-penalized Cox regression methods and determine similarity and differences among the models in the prediction of survival time and short-term survival in high dimensional survival datasets generated by varying censoring rates. Material and Methods: Gene expression survival datasets containing n=200 units and p=1000 gene expression levels whose correlation levels were changing between -0.7 and 0.7 were randomly generated. Simulated datasets were then randomly divided into training and test sets in a 70:30 ratio. Extreme learning machines-based survival, supervised principal components, and L2-penalized Cox regression models were trained in a training set. At the end of the 1000 times repetitive simulation, Harrell's concordance index value, integrated Brier score, sensitivity, specificity, accuracy rates, and negative predictive value, positive predictive value, the area under precision-recall, area under the curve, F1 score, Cohen's kappa coefficient, and Matthew's correlation coefficient were calculated to reveal the performances of the methods. Results: When the simulation results were examined, it was determined that the survival models' performances were close to each other. It was also observed that the performances of survival models concerning the prediction of both survival time and short-term survival tend to decrease by increasing the censoring rate. According to the applied hierarchical clustering analysis, it was determined that the methods that perform close to each other according to the varying censoring rates were in the same cluster. It was noted that in all scenarios, an extreme learning machine Cox model with likelihood-based boosting and L2-penalized Cox methods were the methods that showed the closest performance to each other. In contrast, an extreme learning machine Cox model with a gradient-based boosting method showed far lower performance than other methods. Conclusion: To conclude, the high rate of censoring in survival data adversely affects the performance of the survival models. Since the performances of the models in the analysis of high-dimensional survival data were close to one another, it was revealed that extreme learning machines-based survival models, which can directly analyze high-dimensional survival data, were useful and can be preferred instead of dimension reduction methods such as supervised principal component analysis, and penalized models. Keywords: Extreme learning machines, Penalized Cox regression model, Simulation, Supervised principal components, Survival