Fig. 3.

Trained random survival forest model is highly accurate. (A) Schematic illustrating the general random survival forest algorithm as well as key hyperparameters. (B) Hyperparameter tuning grid and interpolated performance surface for minimum terminal node size and number of variables used per split. Lower values of OOS error (better) are colored in orange, and higher values are colored in blue. White dots represent each tested grid point, with the minimal training error identified by an X (node size = 1; variables per split = 32). (C) Cumulative OOS performance versus number of trees up to 3000 in the random survival forest using final nodesize and mtry parameter settings. (D) Aggregate predicted EFS and 95% confidence intervals for training cohort patients who experienced events (red) and did not (teal), with OOS error rate (1 – C-index) of 17.93%. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)