Title: SurvPFN: Advancing Foundation Models for Survival Analysis

Abstract: While Tabular Foundation Models (TFMs) have achieved significant advancements in conventional classification and regression tasks, they have largely failed to address time-to-event survival prediction. This gap exists because survival analysis requires handling censored data, a complexity that standard TFMs cannot natively process. Consequently, applying standard TFMs to these tasks results in biased and imprecise forecasts, rendering them ineffective for practical, real-world scenarios. To address this critical deficiency, we introduce \texttt{SurvPFN}, a prior-data fitted network (PFN) specifically designed for survival prediction. Through pretraining on millions of synthetic survival datasets, \texttt{SurvPFN} acquires the ability to model survival via distributional regression that properly accounts for censoring. The model operates through three key mechanisms: (1) the generation of data featuring Weibull-distributed event times alongside a non-informative censoring process; (2) the incorporation of indicators for censored events; and (3) the optimization of a censored negative log-likelihood. Evaluated on SurvSet—a benchmark comprising real-world survival tasks—\texttt{SurvPFN} demonstrates strong competitiveness against both classical and deep learning survival baselines. Notably, it achieves this performance without requiring per-dataset fine-tuning, specialized survival architectures, or extensive feature engineering. Our findings suggest that survival analysis can be effectively framed as a continuous-time distributional regression problem utilizing censored loss functions, thereby harnessing the capabilities of PFNs for time-to-event predictions.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC