Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SurvivalPFN, a prior-data fitted network pretrained on synthetic identifiable right-censored DGPs that amortizes Bayesian posterior predictive survival prediction via in-context learning in a single forward pass.
Problem
Survival analysis requires choosing among many specialized estimators with distinct assumptions and tuning; per-dataset Bayesian inference for censored data is expensive and intractable for flexible models.
Approach
Pretrain a transformer PFN on a broad prior over identifiable right-censored survival DGPs (conditional independent censoring) so that, given an observed censored dataset as context, it outputs posterior predictive event-time distributions for query covariates in one forward pass, with query indicators selecting event or censoring targets.
Key results
- SurvivalPFN achieves the best median rank across 61 datasets and 5 metrics among 21 survival models.
- SurvivalPFN is among the leading methods on IBS, MAE, and Log-Rank, while RSF, CWGB, and DeepSurv rank higher on CI and D-calibration.
- SurvivalPFN is only modestly slower than CoxNet, the fastest baseline, in total training-plus-inference time.
Surprising findings
- Generic tabular foundation models trained only on uncensored instances are outperformed by SurvivalPFN, suggesting survival-specific pretraining is needed.
Limitations
- Relies on conditional independent censoring for identifiability; cannot handle dependent censoring.
- Reduced relative performance on larger tables due to PFN context-size scalability limits.
- Concurrent TFM-based survival methods (SIC) not directly compared since weights/code unavailable.
Experimental setup
Datasets: SurvSet textbook datasets software-package datasets recent-publication datasets
Benchmarks: SurvSet-derived 61-dataset benchmark
Metrics: IBS concordance index D-calibration median-time MAE log-rank
Baselines: CoxPH CoxNet cSVR GB CWGB RSF DeepHit DeepSurv StaticSurvivalTFM DSM
Builds on: prior-data fitted networks TabDPT CausalPFN SurvSet
Closest papers by meaning
- Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees (AAAI 2024 · similarity 0.87)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.87)
- Toward Conditional Distribution Calibration in Survival Prediction (NeurIPS 2024 · similarity 0.87)
- Survival Analysis via Density Estimation (ICML 2025 · similarity 0.86)
- NeuralSurv: Deep Survival Analysis with Bayesian Uncertainty Quantification (NeurIPS 2025 · similarity 0.86)
- Gradient-based Explanations for Deep Learning Survival Models (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.