Survival Analysis with triageR

library(triageR)
library(survival)

Overview

Not every clinical outcome is binary. Many research questions are about time to an event, time to death, relapse, readmission, or disease progression — where some patients are followed for the full study and others are “censored” (the event hasn’t happened by the time we stop observing them, or they leave the study early).

This vignette walks through triageR’s survival analysis workflow using the classic lung dataset (from the survival package): time to death for patients with advanced lung cancer.

1. Prepare the data

The lung dataset codes its event status as 1 (censored) / 2 (death), a common convention in older survival datasets, but triageR expects standard 0/1 coding (0 = censored, 1 = event occurred).

lung_clean <- lung
lung_clean$status <- lung_clean$status - 1
lung_clean <- lung_clean[stats::complete.cases(lung_clean), ]

head(lung_clean[, c("time", "status", "age", "sex", "ph.karno")])
#>   time status age sex ph.karno
#> 2  455      1  68   1       90
#> 4  210      1  57   1       90
#> 6 1022      0  74   1       50
#> 7  310      1  68   2       70
#> 8  361      1  71   2       60
#> 9  218      1  53   1       70

2. Fit a survival model

triageR supports two survival engines through the same interface used for classification models: Cox Proportional Hazards (the clinical standard, via the survival engine) and Random Survival Forest (via the aorsf engine).

Unlike tr_fit(), survival models are fit with tr_fit_survival(), which expects a time_col and event_col instead of a single outcome column — reflecting the different shape of time-to-event data.

model_cox <- tr_fit_survival(
  lung_clean,
  time_col = "time",
  event_col = "status",
  engine = "cox_ph"
)
#> Survival model fitted successfully using engine: cox_ph

model_rf <- tr_fit_survival(
  lung_clean,
  time_col = "time",
  event_col = "status",
  engine = "survival_rf"
)
#> Survival model fitted successfully using engine: survival_rf

3. Validate the model

Survival models are validated with tr_validate_survival(), which computes the concordance index (C-index) — survival analysis’s equivalent of AUC. A C-index of 0.5 means the model is no better than chance at ranking which patient will experience the event sooner; 1.0 means perfect discrimination.

val_cox <- tr_validate_survival(model_cox, newdata = lung_clean)
#> Survival validation (training_data):
#> C-index: 0.648 (SE: 0.03)
#> N = 167, events = 120
#> 
#> Interpretation: C-index of 0.5 = no better than chance; 1.0 = perfect discrimination between patients who experience the event sooner vs. later.

val_rf <- tr_validate_survival(model_rf, newdata = lung_clean)
#> Survival validation (training_data):
#> C-index: 0.839 (SE: 0.018)
#> N = 167, events = 120
#> 
#> Interpretation: C-index of 0.5 = no better than chance; 1.0 = perfect discrimination between patients who experience the event sooner vs. later.

In this example, the Random Survival Forest achieves a notably higher C-index than Cox Proportional Hazards, suggesting the relationship between predictors and survival time may not be fully linear on the log hazard scale, exactly the kind of finding that motivates comparing more than one modelling approach.

4. Automated pipeline review

tr_agent_review() works with survival models too, adapting its checks to the time-to-event context: instead of class imbalance, it checks the event rate (the proportion of patients who experienced the event rather than being censored), and events-per-variable is calculated using the number of observed events rather than total sample size.

review_cox <- tr_agent_review(lung_clean, model_cox, use_agent = FALSE)
#> 
#> --- triageR Pipeline Review ---
#> 
#> No major issues flagged.

5. Generate a TRIPOD+AI-aligned report

tr_tripod_report() supports survival models via the model_type argument. The report adapts automatically: instead of a confusion matrix and calibration plot, it reports the C-index and its standard error.

tr_tripod_report(
  model = model_cox,
  model_type = "survival",
  validation = val_cox,
  review = review_cox,
  output_file = file.path(tempdir(), "lung_survival_report"),
  format = "html"
)

Note: automated sensitivity analysis (tr_sensitivity()) is currently only supported for classification models, not survival models.

Summary

This vignette covered triageR’s survival analysis workflow: preparing time-to-event data, fitting Cox Proportional Hazards and Random Survival Forest models through a consistent interface, validating with the concordance index, running an automated pipeline review adapted for censored data, and generating a TRIPOD+AI-aligned report, all using a real, widely-used clinical dataset.