Train and Evaluate a Survival Model

This function trains and evaluates a single survival model using the tidymodels framework. It supports a range of survival model types, including Cox, penalized Cox, AFT, random forests, bagged trees, and gradient boosting models.

Usage

compute_ml_survival(
  df_train,
  df_test,
  outcome_col,
  event_col = NULL,
  model,
  models_hyperparameters
)

Arguments

df_train

A data frame containing the training data, including survival time, event indicator, and predictor variables.

df_test

A data frame containing the test data with the same structure and columns as df_train.

outcome_col

Character string specifying the name of the survival time column.

event_col

Character string specifying the name of the event indicator column (1 = event occurred, 0 = censored).

model

Character string specifying which survival model to train. Supported values include:

• "cox_ph_survival" – Cox proportional hazards model (survival)

• "proportional_hazards_glmnet" – Penalized Cox model (LASSO/Elastic Net)

• "survreg_flexsurv" – Parametric AFT model (flexsurv)

• "rand_forest_partykit" – Random survival forest (ctree engine)

• "rand_forest_aorsf" – Oblique random survival forest

• "decision_tree_partykit" – Single survival tree

• "bag_tree_rpart" – Bagged survival trees

• "boost_tree_mboost" – Gradient boosting for survival data

models_hyperparameters

A list containing model hyperparameter values to apply. Typically created from a tuning grid or optimization step, e.g.: list(list(trees = 500, min_n = 10)). If NULL, default parameters are used.

Value

A tibble with two columns:

model

The model name as a character string.

c_index

Numeric value of the computed C-index on the test data.

Details

The function automatically applies user-specified hyperparameters, standardizes predictions from different engines, and evaluates model performance using the Concordance Index (C-index) as the primary metric. It ensures a consistent risk-score direction across models (higher values indicate higher risk).

The function uses the parsnip interface from tidymodels to define, train, and evaluate models.

Workflow:

1. Defines the model specification (parsnip::model_spec).

2. Optionally applies user-specified hyperparameters via parsnip::set_args().

3. Constructs a survival formula of the form Surv(time, event) ~ ..

4. Fits the model using parsnip::fit() on the training data.

5. Evaluates model performance using predict_and_evaluate_survival(), which computes the C-index.

See also

predict_and_evaluate_survival()