Train a Stacking Prognostic Model

Implements a Stacking ensemble for prognostic models. It trains multiple base models and uses their predictions to train a meta-model.

Usage

stacking_pro(
  results_all_models,
  data,
  meta_model_name,
  top = 3,
  tune_meta = FALSE,
  time_unit = "day",
  years_to_evaluate = c(1, 3, 5),
  seed = 789
)

Arguments

results_all_models

A list of results from models_pro(), containing trained base model objects and their evaluation metrics.

data

A data frame for training the meta-model. The first column must be ID, second event status (0/1), third time, and subsequent columns features.

meta_model_name

A character string, the name of the meta-model to use (e.g., "lasso_pro", "gbm_pro"). This model must be registered.

top

An integer, the number of top-performing base models (ranked by C-index) to select for the stacking ensemble.

tune_meta

Logical, whether to enable tuning for the meta-model.

time_unit

A character string, the unit of time in the third column of data.

years_to_evaluate

A numeric vector of specific years at which to calculate time-dependent AUROC for evaluation.

seed

An integer, for reproducibility.

Value

A list containing the model_object, sample_score, and evaluation_metrics.

See also

models_pro, evaluate_model_pro

Examples

# \donttest{
# NOTE: This example requires the 'train_pro' dataset.
if (requireNamespace("E2E", quietly = TRUE) &&
"train_pro" %in% utils::data(package = "E2E")$results[,3]) {
  data(train_pro, package = "E2E")
  initialize_modeling_system_pro()

  # First, generate results for base models
  base_model_results <- models_pro(data = train_pro, model = c("lasso_pro", "rsf_pro"))

  # Then, create the stacking ensemble
  stacking_lasso_results <- stacking_pro(
    results_all_models = base_model_results,
    data = train_pro,
    meta_model_name = "lasso_pro",
    top = 3,
    years_to_evaluate = c(1, 3)
  )
  print_model_summary_pro("Stacking (Lasso)", stacking_lasso_results)
}
#> Prognosis modeling system already initialized.
#> Running model: lasso_pro
#> Running model: rsf_pro
#> Running Stacking model: Stacking_pro (meta: lasso_pro)
#> 
#> --- Stacking (Lasso) Prognosis Model (on Training Data) Metrics ---
#> Ensemble Type: Stacking (Meta: lasso_pro, Base models used: rsf_pro, lasso_pro)
#> C-index: 0.8814
#> Time-dependent AUROC (years 1, 3): 0.7389, 0.8173
#> Average Time-dependent AUROC: 0.7781
#> KM Group HR (High vs Low): 19.5197 (p-value: 3.159e-18, Cutoff: 18.0175)
#> --------------------------------------------------
# }