Title: Maintaining Vitality: Unrestricted Survival Analysis Using Tabular Foundation Models

Abstract: Survival Analysis (SA) constitutes a statistical approach designed to model the duration preceding the occurrence of a specific event of interest. While this framework is extensively applied across various sectors, such as churn prediction and healthcare, its practical implementation is often hindered by right-censoring, a scenario where the event time is only partially observed. Recently, Tabular Foundation Models (TFMs) have garnered considerable attention because they can execute prediction tasks in a single forward pass, eliminating the need for dataset-specific parameter optimization. However, applying these models to time-to-event prediction remains challenging due to the issue of right censoring. This study introduces a training-free technique for survival regression that utilizes TFMs to simultaneously forecast event times and iteratively impute censored data. By employing a TFM to establish an Accelerated Failure Time (AFT) model, our approach requires training only a single scalar parameter. Furthermore, extending from the Buckley-James estimator, we propose a non-parametric in-context estimator specifically designed for right-censored data. Evaluations on standard survival analysis benchmarks demonstrate that our method performs competitively against various trained parametric and semi-parametric survival regression models, including parametric AFT models and Cox regression.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC