Predict and Evaluate Survival Model Performance

This function generates predictions from a fitted survival model and evaluates its performance using the Concordance Index (C-index). It automatically handles different prediction output types supported by various survival modeling engines and standardizes predictions into a comparable numeric format.

Usage

predict_and_evaluate_survival(
  model_fit,
  data,
  outcome_col = "time",
  event_col = "event"
)

Arguments

model_fit

A fitted survival model object (typically created via a parsnip or workflow model specification).

data

A data frame containing the predictors and survival outcome variables used for prediction and evaluation.

outcome_col

Character string specifying the name of the survival time column in data. Default is "time".

event_col

Character string specifying the name of the event indicator column in data. Default is "event".

Value

A list with two elements:

preds

A tibble containing standardized numeric predictions.

c_index

Numeric value representing the computed C-index.

Details

The function attempts to generate predictions using multiple possible prediction types, depending on what the model supports:

• "linear_pred" — Linear predictor or log hazard (e.g., Cox models), where higher values indicate higher risk.

• "time" — Expected survival time (e.g., AFT models), where higher values imply longer survival (automatically reversed).

• "survival" — Survival probability at a specific evaluation time, where higher values indicate better survival (automatically reversed).

It standardizes the prediction output into a tibble with one numeric column .pred, ensuring consistency across engines. The function then computes the Concordance Index (C-index) using yardstick::concordance_survival_vec() to assess how well the model ranks predicted survival relative to actual outcomes.

See also

yardstick::concordance_survival_vec(), aggregate_results()

Examples

if (FALSE) { # \dontrun{
fitted_model <- parsnip::fit(
  parsnip::proportional_hazards(engine = "survival"),
  Surv(time, event) ~ .,
  data = lung
)
results <- predict_and_evaluate_survival(fitted_model, lung)
results$c_index
} # }