survcompare

The R package survcompare,  developed by Dr Diana Shamsutdinova, investigates the complexity of survival data and performs nested cross-validation of the Cox-PH, Survival Random Forest, DeepHit models, as well as their ensembles. 

The package tests the significance of the performance differences between the linear Cox-PH model and its ML alternatives, Survival Random Forests and DeepHit. The presence of the complex data relationships and their contribution to the predictive accuracy is inferred from the performance difference.

• Package link: https://github.com/cran/survcompare

• Vignette:  Survcompare_application 

• Extention to DeepHit: https://github.com/dianashams/survcompare/tree/DeepHit

• Corresponding paper:  Shamsutdinova, D., Stamate, D., & Stahl, D. (2025). Balancing accuracy and Interpretability: An R package assessing complex relationships beyond the Cox model and applications to clinical prediction. International Journal of Medical Informatics, 194, 105700.  https://www.sciencedirect.com/science/article/pii/S1386505624003630

• Getting started:  install the package from CRAN as install.packages("survcompare"), or from its github directory by running the devtools::install_github("dianashams/survcompare") command. The main function to use is survcompare(data, predictors). The data should be in a form of a data frame, with "time" and "event" columns defining the survival outcome. A list of column names corresponding to the predictors to be used should also be supplied. Missing data can be handled by missForest by setting "impute" to 1 in survcompare. 

• Video presentation: https://www.youtube.com/watch?v=1Z8C0pAi_Cs&t=807s