Skip to content

Method overview

SurvFM separates RMST target construction from the tabular prediction engine. One outer training fold and one restriction horizon define the complete workflow.

  1. Observed data. Supply baseline covariates, observed follow-up and a binary first-event indicator.
  2. Horizon. Use a prespecified decision horizon, or derive a training-fold event-time quantile for a generic benchmark.
  3. Target construction. Estimate marginal Kaplan–Meier RMST and construct exact delete-one pseudo-observations.
  4. Preprocessing. Fit numeric and categorical transformations only on the training fold.
  5. Regression. Supply the same pseudo-RMST context to an unchanged tabular backbone.
  6. Prediction. Return bounded RMST at the selected horizon; use its negative only when a risk-oriented score is required.

The method is estimand-focused: changing the backbone does not change the target definition. Conversely, changing τ changes the prediction question and usually requires constructing a new target and fitting a horizon-specific model.

Minimal mathematical map

Right-censored training fold Kaplan–Meier RMST at τ Jackknife pseudo-RMST labels Tabular regression backbone Held-out RMST prediction

The pseudo-RMST derivation gives the statistical construction; the shared regression interface shows how it is supplied to different backbones.