Figure 5.

Survival models performance. (A) Boxplot of the concordance index (c-index) in 100-times repeated 5-fold cross validation, depending on the survival algorithm and the predictor variables initially entered in the modeling. Other abbreviations: Step. Cox, stepwise Cox regression; GBM, gradient boosted machine; LASSO, least absolute shrinkage and selection operator Cox regression; RSF, random survival model. (B) Brier score (with 95% confidence internal) in 100-times repeated 5-fold cross validation as a function of time (i.e., prediction error curve), between immunotherapy beginning and 2 years, for the benchmark model (stepwise Cox regression on the clinicopatholologic [c.p.] input), the best-performing shallow-learning model (GBM on full input), and the best-performing deep-learning model (Deepsurv on uncorrelated input).