A Comparison of Cox Model and Machine Learning Techniques in the High-Dimensional Survival Data

Predicting certain events of interest that will occur at future time points is the primary objective of survival analysis. Machine Learning (ML) algorithms have been extensively used in the field of survival analysis especially cancer prognosis studies in recent years. ML algorithms like Support Vector Machines (SVM) and Random Survival Forests (RSF) are extensively used for the censored survival data in the model building. However, ML algorithms and the traditional Cox regression approach result in better prediction performance but still challenging to decide the choices of the approach. Also, the prediction performance of both approaches is limited in high-dimensional data. Therefore, this study mainly aims to investigate the variable selection method called regularized regression like Least Absolute Shrinkage and Selection Operator (LASSO), ridge, and elastic net to improve the prediction performance for the different training fractions using a high-dimensional breast cancer dataset GSE7390 with survival endpoints from the Gene Expression Omnibus database. The results of this study reveal that LASSO performs better than the other regularized regression methods as its C-index is higher with minimum prediction error and holds high prediction accuracy for both SVM and RSF for different training fractions. While comparing the LASSO-SVM and LASSO-RSF with LASSO-COX, the Cox model resulted in high prediction accuracy, and satisfactory C-index with minimum prediction error for different training fractions.