Is it time to use machine learning survival algorithms for survival and risk factors prediction instead of Cox proportional hazard regression? A comparative population-based study

Purpose Applying machine learning in medical statistics offers more accurate prediction models. In this paper, we aimed to compare the performance of the Cox Proportional Hazard model (CPH), Classification and Regression Trees (CART), and Random Survival Forest (RSF) in short-, and long-term prediction in glioblastoma patients. Methods We extracted glioblastoma cancer data from the Surveillance, Epidemiology, and End Results database (SEER). We used the CPH, CART, and RSF for the prediction of 1- to 10-year survival probabilities. The Brier Score for each duration was calculated, and the model with the least score was considered the most accurate. Results The cohort included 26473 glioblastoma patients divided into two groups: training (n = 18538) and validation set (n = 7935). The average survival duration was seven months. For the short- and long-term predictions, RSF was the best algorithm followed by CPH and CART. Conclusion For big data, RSF was found to have the highest accuracy and best performance. Using an accurate statistical model for survival prediction and prognostic factors determination will help the care of cancer patients. However, more developments of the R packages are needed to allow more illustrations of the effect of each covariate on the survival probability.