Skip to content

SurvFM

Survival prediction with reusable tabular backbones

Research software. SurvFM is intended for methodological research and retrospective evaluation. It has not been validated for clinical decision-making.

SurvFM is a model-agnostic framework that adapts tabular prediction models to right-censored survival outcomes. It constructs jackknife pseudo-observation targets for restricted mean survival time (RMST), then supplies the same horizon-specific regression task to interchangeable tabular backbones.

Install SurvFM View the source

SurvFM framework

1. Define the horizon
Use a clinically relevant restriction time with adequate follow-up support, or a training-fold event-time quantile for benchmarking.
2. Construct targets
Convert censored training outcomes into exact jackknife pseudo-RMST labels.
3. Fit a backbone
Use TabPFN by default, select another registered backbone, or supply a custom regressor.
from survfm import fit_survfm

model = fit_survfm(
    X_train,
    time_train,
    event_train,
    backbone="tabpfn",
    tau=365.0,
)

predicted_rmst = model.predict(X_test)
risk_score = model.predict_risk(X_test)

Scope

  • Static baseline covariates and a binary first-event endpoint.
  • Horizon-specific predictions targeting covariate-conditional RMST.
  • Exact, clipped jackknife pseudo-observations matching the manuscript workflow.
  • Train-only mixed-data preprocessing for pandas DataFrames.
  • Local and hosted backbones through one estimator API.

What a prediction means

A SurvFM prediction targets the expected event-free time accumulated up to the chosen horizon among observations with similar baseline characteristics. It is expressed in the original time units and is not an individual's predicted event date. Negative predicted RMST can be used when a higher-is-riskier score is needed for concordance evaluation.

Naming

  • SurvFM names the model-agnostic framework.
  • Pseudo-RMST target construction names the statistical mechanism used to transform censored follow-up into regression labels.
  • SurvFM-TabPFN, SurvFM-TabICL and related labels identify SurvFM paired with a particular tabular backbone.