cs.LGMar 31, 2026

Survival In-Context: Amortized Bayesian Survival Analysis via Prior-Fitted Networks

Authors: Dmitrii Seletkov, Paul Hager, Georgios Kaissis, Rickmer Braren, Daniel Rueckert, Raphael Rehms

Organizations: Institute of Diagnostic and Interventional Radiology, Technical University of Munich, Germany · Chair for AI in Healthcare and Medicine, Technical University of Munich, Germany · Hasso Plattner Institute for Digital Engineering, University of Potsdam, Germany · University Hospital Hamburg-Eppendorf, Germany · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Survival analysis is crucial for many medical applications, but remains challenging for modern machine learning due to limited data, censoring, and the heterogeneity of tabular covariates. While the prior-fitted paradigm, which relies on pretraining models on large collections of synthetic datasets, has recently facilitated tabular foundation models for classification and regression, its suitability for time-to-event modeling remains unclear. We propose a flexible survival data generation framework that defines a rich survival prior with explicit control over covariates and time-event distributions. Building on this prior, we introduce Survival In-Context (SIC), a prior-fitted in-context learning model for survival analysis that is pretrained exclusively on synthetic data. SIC is trained to approximate Bayesian posterior predictive inference under the synthetic survival prior, enabling individualized survival prediction in a single forward pass, requiring no task-specific training or hyperparameter tuning. Across a broad evaluation on real-world survival datasets, SIC achieves competitive or superior performance compared to classical and deep survival models, particularly in small and medium-sized data regimes, highlighting the promise of a prior-fitted paradigm for survival analysis. The code and pretrained models will be made available upon publication.

Figures & tables

Figure 1: Survival In-context (SIC) method contains two stages. (I) Synthetic Data Generation using Structural Causal Models (SCMs) for the generation of (I.A) covariates X and the associated label nodes η\eta , followed by (I.B) two proposed survival priors to generate time-event labels: continuous (I.B.1) with the survival function parametrization H0−1(∗,α,β)H_{0}^{-1}(*,\alpha,\beta) and discrete (I.B.2) with Monotonic Positive Network [ 66 ] fωf_{\omega} . (II) In-context learning on synthetic data with a specialized time-event embedding and survival head and loss. Created in https://BioRender.com .
Figure 2: Visualization of a batch of 512 generated datasets using SCM η1=X1+X2⋅X3\eta_{1}=X_{1}+X_{2}\cdot X_{3} and η2=X1⋅X2\eta_{2}=X_{1}\cdot X_{2} for continuous (upper) and discrete (lower) survival priors: time-dependent C-index estimated from fit CoxPH models (left), Kaplan–Meier curves (middle), and the underlying event survival curves on the normalized time scale. The proposed synthetic data generation methods demonstrate the ability to generate diverse survival curve patterns and outcome complexity.
Figure 3: Time-dependent C-index performance (mean and std) across 5 folds on 12 real-world datasets. The proposed Survival In-context (SIC) method achieves comparable performance to the baselines CoxPH [ 16 ] , DeepHit [ 45 ] , DeepSurv [ 36 ] , XGB [ 14 ] , RSF [ 33 ] , ConSurv [ 46 ] , TabICL [ 63 ] + Stacking [ 17 ] . SIC does not require hyperparameter tuning, whereas all baselines are tuned with 100 trials. * indicates p-value << 0.05 for the two-sided t-test of SIC compared to the baselines.
Figure 4: Dataset-wise method ranks (best = 1) across 12 datasets, ordered by dataset size. The proposed Survival In-context (SIC) method shows the highest mean ( ⋆\star ) and median ( △)\triangle) rank compared to the baselines CoxPH [ 16 ] , DeepHit [ 45 ] , DeepSurv [ 36 ] , XGB [ 14 ] , RSF [ 33 ] , ConSurv [ 46 ] , TabICL [ 63 ] + Stacking [ 17 ] .
Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

DistributionH0−1(y)H_{0}^{-1}(y)
Weibullα y1/β; α >0, β>0\displaystyle\alpha\,y^{1/\beta}\text{; }\alpha\ >0\text{, }\beta>0
Gompertz1α ln⁡ ⁣(1+αβ y); α >0, β≠0\displaystyle\frac{1}{\alpha}\,\ln\!\left(1+\frac{\alpha}{\beta}\,y\right)\text{; }\alpha\ >0\text{, }\beta\neq 0
Lognormalexp⁡ ⁣(α+β Φ−1(1−e−y)); α >0\displaystyle\exp\!\Big(\alpha+\beta\,\Phi^{-1}(1-e^{-y})\Big)\text{; }\alpha\ >0
Log-logisticα (ey−1)1/β; α >0, β>0\displaystyle\alpha\,\big(e^{y}-1\big)^{1/\beta}\text{; }\alpha\ >0\text{, }\beta>0
Birnbaum–Saundersα[12(β Φ−1(1−e−y)+(β Φ−1(1−e−y))2+4)]2; α >0, β>0\displaystyle\alpha\left[\tfrac{1}{2}\Big(\beta\,\Phi^{-1}(1-e^{-y})+\sqrt{(\beta\,\Phi^{-1}(1-e^{-y}))^{2}+4}\Big)\right]^{2}\text{; }\alpha\ >0\text{, }\beta>0
Appendix
Table 1: Inverse cumulative hazard H0−1(y)H_{0}^{-1}(y) for baseline survival distributions.
Figure 5: Visualization of control parameters for a discrete survival prior: survival (left) and Kaplan-Meier (right) curves for a batch of 100 datasets.
Dataset#samples#featuresEvent rate 
VETERAN137693.4% 
LUNG2881072.4% 
PBC3122044.9% 
WHAS5005001443.0% 
GBSG2686943.6% 
AIDS1,151118.3% 
Appendix
Table 2: Dataset statistics. UNOS and SEER datasets have restricted access and exceed the pretraining prior (sample size) of SIC.
ModelHyperparameterDomain
Cox [ 16 ]l1_ratio[0.0, 1.0]
 penalizer[1e-6, 1]
DeepSurv [ 36 ]optimizer{adam, sgd}
 weight_decay[1e-6, 1e-2]
 momentum[0.0, 0.9]
 lr[1e-5, 1e-2]
Appendix
Table 3: Summary of hyperparameter search spaces
DatasetSICCoxPHDeepSurvDeepHitXGBRSFTabICL + StackingConsurv
VETERAN0.720 (0.020)0.724 (0.039)0.687 (0.044)0.632 (0.105)0.669 ∗ (0.031)0.686 (0.044)0.660 (0.076)0.586 ∗ (0.056)
LUNG0.615 (0.060)0.615 (0.053)0.577 (0.106)0.514 (0.075)0.562 ∗ (0.074)0.587 ∗ (0.072)0.565 (0.032)0.606 (0.027)
PBC0.819 (0.067)0.790 (0.049)0.793 (0.060)0.812 (0.064)0.814 (0.056)0.808 (0.062)0.747 ∗ (0.072)0.790 (0.062)
WHAS5000.757 (0.026)0.766 (0.037)0.754 (0.039)0.765 (0.026)0.758 (0.043)0.773 (0.009)0.665 ∗ (0.036)0.747 (0.041)
GBSG20.691 (0.040)0.680 (0.024)0.656 (0.076)0.667 (0.050)0.671 (0.039)0.681 (0.037)0.558 ∗ (0.034)0.658 ∗ (0.035)
AIDS0.743 (0.046)0.743 (0.030)0.662 (0.072)0.685 (0.073)0.735 (0.042)0.715 (0.048)0.670 (0.057)0.755 (0.026)
Appendix
Table 4: Time-dependent C-index (mean (std) across 5 folds). ∗ indicates paired t-test significance vs. SIC ( α=0.05\alpha=0.05 ).
Figure 6: Time-dependent C-index of SIC framework with different Survival Priors : Discrete-only, Continuous-only, and Mix of both (main option used in evaluation).
DatasetSICCoxPHDeepSurvDeepHitXGBRSFTabICL + StackingConsurv
UNOS0.571 (0.004)0.589 ∗ (0.005)0.595 ∗ (0.005)0.611 ∗ (0.006)0.599 ∗ (0.006)–0.529 ∗ (0.013)0.595 ∗ (0.004)
SEER0.840 (0.004)0.820 ∗ (0.003)0.848 ∗ (0.005)0.853 ∗ (0.004)0.867 ∗ (0.003)–0.804 ∗ (0.004)0.850 ∗ (0.004)
Appendix
Table 5: Time-dependent C-index (mean (std) across 5 folds). ∗ indicates paired t-test significance vs. SIC ( α=0.05\alpha=0.05 ). Consurv hyperparameter search is reduced to 20 trials compared to the other baselines due to time constraints. – indicates failed experiments (RSF outputs OOM error using sksurv [ 60 ] for larger datasets).
Figure 7: Time-dependent C-index of SIC framework with different discretization: 10 (main option used in evaluation), 50, and 100 bins.
Figure 8: Time-dependent C-index of SIC framework with different Survival Heads : DeepHit (main option used in evaluation), DeepSurv [ 36 ] and DeepEH [ 76 ] . The original implementation of the DeepEH method outputs NaN values for FLCHAIN dataset. This should not be considered for comparison.
Figure 9: D-Calibration performance (number of misscalibrated folds per dataset) across 5 folds on 12 real-world datasets. The Survival In-context (SIC) is a proposed method. SIC-PostCal is the SIC calibrated with CiPOT [ 62 ] . The baselines: CoxPH [ 16 ] , DeepHit [ 45 ] , DeepSurv [ 36 ] , XGB [ 14 ] , RSF [ 33 ] , ConSurv [ 46 ] , TabICL [ 63 ] + Stacking [ 17 ] . SIC does not require hyperparameter tuning, whereas all baselines are tuned with 100 trials. * indicates p-value << 0.05 for the two-sided t-test of SIC compared to the baselines.
Figure 10: Integrated Brier Score performance (mean and std) across 5 folds on 12 real-world datasets. The Survival In-context (SIC) is a proposed method. SIC-PostCal is the SIC calibrated with CiPOT [ 62 ] . The baselines: CoxPH [ 16 ] , DeepHit [ 45 ] , DeepSurv [ 36 ] , XGB [ 14 ] , RSF [ 33 ] , ConSurv [ 46 ] , TabICL [ 63 ] + Stacking [ 17 ] . SIC does not require hyperparameter tuning, whereas all baselines are tuned with 100 trials. * indicates p-value << 0.05 for the two-sided t-test of SIC compared to the baselines.
DatasetSICSIC-PostCalDelta mean (SIC-PostCal - SIC)
VETERAN0.720 ±\pm 0.0200.721 ±\pm 0.020+0.001
LUNG0.615 ±\pm 0.0600.615 ±\pm 0.0570.000
PBC0.819 ±\pm 0.0670.821 ±\pm 0.065+0.002
WHAS5000.757 ±\pm 0.0260.757 ±\pm 0.0300.000
GBSG20.691 ±\pm 0.0400.690 ±\pm 0.040-0.001
AIDS0.743 ±\pm 0.0460.707 ±\pm 0.092-0.036
Appendix
Table 6: The difference in time-dependent C-intex performance between the proposed Survival In-context (SIC) and SIC after CiPOT [ 62 ] post-calibration (SIC-PostCal).
More in this topic

Survival Analysis

Survival analysis models when events such as disease progression, failure, or death occur, accounting for incomplete follow-up and censoring.

See all 106 papers

Explore similar work

CardsList
May 15, 2026

SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference

Survival analysis provides a powerful statistical framework for modeling time-to-event outcomes in the presence of censoring. However, selecting an appropriate estimator from the many specialized survival approaches often requires substantial methodological and domain expertise. We introduce SurvivalPFN, a prior-data fitted network that amortizes Bayesian inference for censored observations through in-context learning. SurvivalPFN is pretrained on a diverse family of synthetic, identifiable, and right-censored data-generating processes, enabling it to amortize survival analysis in a single forward pass during inference. As a result, the model adapts to the effective complexity of each dataset without task-specific training or hyperparameter tuning, avoids restrictive parametric assumptions, and produces calibrated survival distributions. In a large-scale benchmark spanning 61 datasets, 21 methods, and 5 evaluation metrics, SurvivalPFN achieves strong predictive performance and often improves upon established survival models. These results suggest that SurvivalPFN offers a principled and practical foundation model for survival analysis, with potential applications in high-impact domains such as healthcare, finance, and engineering (https://github.com/rgklab/SurvivalPFN).

Shi-ang Qi, Vahid Balazadeh, Michael Cooper +2

Vector Institute · University of Toronto · University of Alberta +1

Amortized InferenceSurvival AnalysisBayesian InferencePrior-Data Fitted NetworksMortality Prediction
AskPDF
Oct 15, 2026

SurvICL: a Foundation Model for in-context Survival Analysis

Tabular foundation models (TFMs) now outperform gradient-boosted trees in classification and regression. Yet in survival analysis, which predicts when an event occurs from censored data, often on small clinical cohorts, we find that no existing TFM beats the half-century-old Cox model. We address this issue by introducing SurvICL, a TFM that predicts survival curves in-context. SurvICL is pretrained on purely synthetic datasets generated from a novel flexible proportional-odds prior that outperforms other prior types in our ablation study. Unlike most other survival models, SurvICL is trained with a plain quantile regression loss, which is possible because its synthetic pre-training data generator can provide uncensored event times. Additionally, we introduce the largest survival benchmark to date spanning 103 datasets. On the 63 datasets held out from all design decisions, SurvICL outperforms 20 tree-based, deep, and foundation-model baselines without any tuning: it wins 82.9% of pairwise comparisons in integrated Brier score, against 71.3% for the runner-up, the Cox model, and is the first model to rank above Cox on small datasets.

Julie Alberge, Linus Bleistein, Judith Abécassis +3

Inria de Saclay · EPFL · Inria de Saclay, :probabl.

Survival AnalysisTabular Foundation Models
AskPDF
Jun 3, 2026

SurvPFN: Towards Foundation Models for Survival Predictions

Tabular foundation models (TFMs) have made rapid progress in standard classification and regression, but time-to-event survival prediction tasks have remained largely untouched. Unlike in standard regression tasks, survival prediction models must account for censored data. Standard TFMs cannot handle natively censored data, leading to biased and inaccurate predictions, making them unsuitable for real-world applications. To overcome this fundamental limitation, we propose SurvPFN, a prior-data fitted network (PFN), for survival prediction tasks. We pretrain SurvPFN on millions of synthetic survival prediction tasks to learn survival via distributional regression that accounts for censored data. SurvPFN works by (1) generating data with Weibull event times and a non-informative censoring mechanism; (2) integrating a censored event indicator; and (3) minimizing a censored negative log-likelihood. On SurvSet, a collection of real-world survival tasks, SurvPFN is highly competitive with classical and deep survival baselines without per-dataset fitting, a survival-specific architecture, or feature engineering. We show that survival can be treated as a continuous-time distributional regression problem with censored loss, unlocking the power of PFNs for time-to-event predictions.

Samuel Böhm, Lennart Purucker, Frank Hutter +1

Institute of Epidemiology and Prevention, Medical Center - University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany · Institute of Epidemiology and Prevention, Medical Center – University of Freiburg, Freiburg, Germany · Department of Computer Science, University of Freiburg, Freiburg, Germany +7

Tabular Foundation ModelsSurvival AnalysisPrior-Data Fitted NetworksMortality Prediction
AskPDF