cs.LGSep 4, 2026

Adaptation Interfaces for In-Context Tabular Foundation Models in Time-to-Event Prediction

Authors: Minh-Khoi Pham, Luca Cotugno, Dan Cernei, Alina Sirbu, Stefano Masi, Giuseppe Prencipe, Alessandro Pingitore, Patrizia Landi, 

Organizations: ADAPT Centre, School of Computing, Dublin City University DCU Glasnevin Campus, Dublin 9, D09 V209, Ireland · Department of Computer Science and Engineering, University of Bologna, Italy · University of Pisa, Italy · Institute of Clinical Physiology, CNR, Pisa, Italy · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Tabular foundation models (TabFMs) achieve strong performance on structured data, particularly for standard classification and regression problems. Yet, extending them to censored time-to-event prediction is challenging because it requires properly handling censoring and event-time dynamics. Building on our prior work, we further link TabFMs with CoxPH and DeepHit and revise the context-resampled training procedure. We evaluate temporal zero-shot reformulation, classification-based fine-tuning, and survival-head adaptation using frozen TabFM backbones on 74 single-risk data sets, and we additionally study 4 competing-risk data sets. Zero-shot inference is effective on smaller single-risk data sets, whereas supervised adaptation becomes increasingly advantageous as data sets scale. Cox provides the most reliably strong interface, especially for Integrated Brier Score (IBS) on larger data sets. DeepHit is relatively stronger for the time-dependent Concordance Index than for IBS, while cause-specific MTLR ranks highest among the TabFM survival heads in the four-data-set competing-risk analysis. Classification fine-tuning becomes more competitive with zero-shot inference as data sets grow but remains weaker for probabilistic prediction. Overall, our results indicate that effective TabFM transfer depends on the data regime and on the statistical structure represented by the chosen adaptation interface. The implementation scripts used for this work are available at https://github.com/kaylode/survival-fm.

Figures & tables

Figure 1: Adaptation and backbone effects across 74 single-risk data sets. Blue compares survival-head adaptation with CE using the same frozen TabFM backbone; orange compares TabFM-based survival models with matched non-pretrained neural baselines. Panel (a) shows changes in CtdC_{\mathrm{td}} and panel (b) changes in IBS, sign-reversed so that positive values indicate improvement.
Figure 2: Size-stratified mean-rank comparison for single-risk survival analysis. Lower rank is better. Results are aggregated across 40 small, 25 medium, and 9 large data sets and compare discrimination using time-dependent CtdC_{td} with probabilistic prediction using IBS.
Figure 3: Mean-rank comparison for competing-risk survival across four data sets. Lower rank is better. The comparison includes the three survival-head interfaces and classical, tree-based, and non-pretrained neural baselines; horizon-wise classification interfaces are excluded from this arm.
Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

IDData setNNFeat.Event rateBinsDomainRegime
1ACATH2,23030.6610Cardiologymedium
2ACTG1,151170.0810Infectious diseasemedium
3AIDS93850.1510Infectious diseasemedium
4AIDS22,810120.6250Infectious diseasemedium
5AML BULL11562830.5710Haematologysmall
6BERGAMASCHI82100.3410Oncologysmall
Appendix
Table 1: Data set characteristics. NN : number of subjects. Event rate: proportion of uncensored observations. Bins: number of adaptive time intervals used by the single-risk discrete-time formulations. SurvSet data sets are collected from Drysdale (2022) .
IDData setNNFeat.Event rateBinsDomainRegime
Data Sets with Competing Risks
1FRAMINGHAM4,273180.42 ( KK =2)100Cardiologylarge
2PBC21,945160.45 ( KK =2)50Hepatologymedium
3SUPPORT2-CR9,105370.68 ( KK =2)30Critical carelarge
4SYNTHETIC30,000120.50 ( KK =2)10Syntheticlarge
 Min1,945120.4210  
Appendix
Table 2: Data set characteristics for data sets with competing risks. NN : number of subjects. Event rate: proportion of uncensored observations. Bins: number of discrete time intervals used by discrete-time models; interval construction depends on the formulation, with cause-specific MTLR using the corresponding adaptive single-risk discretization and joint DeepHit using equal-width intervals over the observed event-time range.
SettingZero-shotClassification FTSurvival-head
Backbone
Gradient updatesNoneNone (frozen)None (frozen)
Trainable headNative classifierClassificationSurvival
Head architecture–1 hidden layer, 64 units, dropout 0.1
Optimization
Optimizer–AdamW; weight decay 10−310^{-3}
Appendix
Table 3: Strategy-level settings used for the TabFM interfaces. A dash denotes a setting that is not applicable.
ModelHyperparameterSearch space
RSFNumber of trees[50,200][50,200] (integer)
Minimum samples to split[5,20][5,20] (integer)
Minimum samples per leaf[5,20][5,20] (integer)
GBSALearning rate[10−2,2×10−1][10^{-2},2\times 10^{-1}] (log-uniform)
Number of estimators[50,200][50,200] (integer)
Maximum depth[2,5][2,5] (integer)
Appendix
Table 4: Hyperparameter search spaces for the tuned non-pretrained baselines. Each model is optimized using 20 Optuna trials within each training fold.
Figure 4: Pairwise model comparisons for CtdC_{\mathrm{td}} across data-size regimes. Cells report data-set-level win scores and mean signed differences. ∗ denotes Benjamini–Hochberg-corrected paired Wilcoxon tests with q≤0.05q\leq 0.05 .
Figure 5: Size-stratified CtdC_{\mathrm{td}} performance gains across TabFM adaptation interfaces. (a) Backbone gains over matched non-pretrained survival models; (b) survival-head gains over same-backbone zero-shot reformulation; (c) survival-head gains over same-backbone classification fine-tuning. ∗ indicates Benjamini–Hochberg-corrected one-sided paired Wilcoxon tests with q≤0.05q\leq 0.05 .
Figure 6: Time-dependent CtdC_{td} as the available labeled training fraction increases from 1% to 100%.
Figure 7: Sensitivity of the classification and discrete-time survival interfaces to the number of temporal bins.
Figure 8: Efficiency–performance trade-offs. Performance versus computational cost on a log scale. Training time for the tuned non-pretrained baselines includes their per-fold Optuna search, which the TabFM variants do not perform, so these values describe the cost of the reported experimental pipelines rather than a direct comparison of optimization speed.
Figure 9: Risk-stratified Kaplan–Meier curves on the COLON data set.
Figure 10: Risk-stratified Kaplan–Meier curves on the DIABETES data set.
AML-BULLAIDS2
ModelCtdC_{td} ↑\uparrowIBS ↓\downarrowAUC ↑\uparrowModelCtdC_{td} ↑\uparrowIBS ↓\downarrowAUC ↑\uparrow
Cox PH0.549_{{\color[rgb]{0.5,0.5,0.5}\pm 0.078}}0.242_{{\color[rgb]{0.5,0.5,0.5}\pm 0.077}}0.519_{{\color[rgb]{0.5,0.5,0.5}\pm 0.132}}Cox PH0.547_{{\color[rgb]{0.5,0.5,0.5}\pm 0.019}}0.138_{{\color[rgb]{0.5,0.5,0.5}\pm 0.015}}0.579_{{\color[rgb]{0.5,0.5,0.5}\pm 0.025}}
RSF\textbf{0.602}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.100}}\textbf{0.184}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.046}}\textbf{0.642}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.134}}RSF0.542_{{\color[rgb]{0.5,0.5,0.5}\pm 0.004}}\textbf{0.136}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.015}}\underline{0.590}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.019}}
GBSA\underline{0.600}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.121}}0.210_{{\color[rgb]{0.5,0.5,0.5}\pm 0.078}}0.593_{{\color[rgb]{0.5,0.5,0.5}\pm 0.167}}GBSA0.458_{{\color[rgb]{0.5,0.5,0.5}\pm 0.053}}0.138_{{\color[rgb]{0.5,0.5,0.5}\pm 0.014}}0.558_{{\color[rgb]{0.5,0.5,0.5}\pm 0.021}}
DeepSurv0.568_{{\color[rgb]{0.5,0.5,0.5}\pm 0.044}}0.206_{{\color[rgb]{0.5,0.5,0.5}\pm 0.044}}0.572_{{\color[rgb]{0.5,0.5,0.5}\pm 0.078}}DeepSurv0.550_{{\color[rgb]{0.5,0.5,0.5}\pm 0.012}}0.139_{{\color[rgb]{0.5,0.5,0.5}\pm 0.013}}0.582_{{\color[rgb]{0.5,0.5,0.5}\pm 0.017}}
MLP-MTLR0.445_{{\color[rgb]{0.5,0.5,0.5}\pm 0.128}}0.217_{{\color[rgb]{0.5,0.5,0.5}\pm 0.055}}0.473_{{\color[rgb]{0.5,0.5,0.5}\pm 0.071}}MLP-MTLR0.537_{{\color[rgb]{0.5,0.5,0.5}\pm 0.031}}0.146_{{\color[rgb]{0.5,0.5,0.5}\pm 0.023}}0.570_{{\color[rgb]{0.5,0.5,0.5}\pm 0.034}}
Appendix
Table 5: Detailed Single-risk datasets Performance Metrics (mean ±\pm std across 5 folds). CtdC_{td} : Antolini time-dependent concordance; AUC: mean dynamic AUC across q25/q50/q75. IBS: Integrated Brier Score.
DatasetModelCtd(1)C_{td}^{(1)} ↑\uparrowIBS (1) ↓\downarrowAUC (1) ↑\uparrowCtd(2)C_{td}^{(2)} ↑\uparrowIBS (2) ↓\downarrowAUC (2) ↑\uparrow
Fram-CRCox PH\textbf{0.744}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.009}}0.177_{{\color[rgb]{0.5,0.5,0.5}\pm 0.007}}\underline{0.639}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.017}}\textbf{0.713}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.020}}0.146_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}0.592_{{\color[rgb]{0.5,0.5,0.5}\pm 0.042}}
 SurvBoost0.730_{{\color[rgb]{0.5,0.5,0.5}\pm 0.004}}0.180_{{\color[rgb]{0.5,0.5,0.5}\pm 0.007}}0.623_{{\color[rgb]{0.5,0.5,0.5}\pm 0.025}}0.701_{{\color[rgb]{0.5,0.5,0.5}\pm 0.018}}0.146_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}0.591_{{\color[rgb]{0.5,0.5,0.5}\pm 0.046}}
 DeepSurv0.691_{{\color[rgb]{0.5,0.5,0.5}\pm 0.057}}0.185_{{\color[rgb]{0.5,0.5,0.5}\pm 0.010}}0.606_{{\color[rgb]{0.5,0.5,0.5}\pm 0.030}}0.568_{{\color[rgb]{0.5,0.5,0.5}\pm 0.138}}0.159_{{\color[rgb]{0.5,0.5,0.5}\pm 0.010}}0.546_{{\color[rgb]{0.5,0.5,0.5}\pm 0.068}}
 MLP-DH0.718_{{\color[rgb]{0.5,0.5,0.5}\pm 0.012}}0.188_{{\color[rgb]{0.5,0.5,0.5}\pm 0.008}}0.621_{{\color[rgb]{0.5,0.5,0.5}\pm 0.010}}0.673_{{\color[rgb]{0.5,0.5,0.5}\pm 0.019}}0.152_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}\textbf{0.597}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.040}}
 DySurv0.720_{{\color[rgb]{0.5,0.5,0.5}\pm 0.010}}0.175_{{\color[rgb]{0.5,0.5,0.5}\pm 0.011}}0.608_{{\color[rgb]{0.5,0.5,0.5}\pm 0.026}}0.679_{{\color[rgb]{0.5,0.5,0.5}\pm 0.029}}0.147_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}0.537_{{\color[rgb]{0.5,0.5,0.5}\pm 0.050}}
 SurvTrace0.728_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}0.176_{{\color[rgb]{0.5,0.5,0.5}\pm 0.010}}0.615_{{\color[rgb]{0.5,0.5,0.5}\pm 0.017}}0.702_{{\color[rgb]{0.5,0.5,0.5}\pm 0.027}}0.144_{{\color[rgb]{0.5,0.5,0.5}\pm 0.005}}0.569_{{\color[rgb]{0.5,0.5,0.5}\pm 0.045}}
Appendix
Table 6: Competing-risk results per cause (mean ±\pm std across 5 folds). Superscript (1)(1) and (2)(2) index the two competing causes. Best value per column within a data set is bold , second best underlined .
More in this topic

Event Time Prediction

Event time prediction estimates when future events will occur from event histories, time series, and contextual information.

See all 44 papers

Explore similar work

CardsList
Jun 10, 2026

Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation

Predicting time-to-event outcomes such as mortality is a fundamental task in clinical decision-making, commonly addressed through survival analysis. While classical statistical and deep learning approaches have been widely studied, they typically require task-specific training and sufficient labeled data. Recent advances in tabular foundation models offer a new paradigm by learning general-purpose representations for structured data. However, their applicability to censored time-to-event prediction in clinical settings remains underexplored, as typical applications are restricted to discrete classification rather than survival analysis tasks. In this work, we propose a lightweight adaptation approach for applying tabular foundation models to clinical survival analysis by directly training a survival-aware head on top of the pretrained representations. We study representative architectures, including TabPFN, TabDPT, and TabICL, and adapt them using a multi-task logistic regression (MTLR) head to model right-censored time-to-event outcomes. We evaluate this approach on a diverse set of public survival benchmarks and two large-scale ICU cohorts, MIMIC-IV and eICU. Our results show that this transfer learning approach achieves competitive or superior performance compared to strong baselines. On MIMIC-IV, TabDPT-FT-MTLR reaches a C-index of 0.856, corresponding to a relative improvement of +1.4% over the best non-FM baseline (DeepSurv, 0.844) and +6.7% over the best zero-shot model (0.802). On eICU, TabICL-FT-MTLR achieves 0.797, yielding gains of +1.7% (DeepSurv, 0.784) and +6.4% (0.749), respectively. These findings highlight the importance of combining pretrained tabular representations with survival-aware objectives and suggest that tabular foundation models provide a practical and effective alternative for clinical survival prediction.

Minh-Khoi Pham, Luca Cotugno, Alina Sirbu +3

ADAPT Centre, Dublin City University, Dublin, Ireland · School of Computing, Dublin City University, Dublin, Ireland · Department of Computer Science and Engineering, University of Bologna, Bologna, Italy

Tabular Foundation ModelsSurvival Analysis
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
Jun 2, 2026

Staying Alive: Uncensored Survival Analysis with Tabular Foundation Models

Survival Analysis (SA) is a statistical framework that models the time span until some event of interest occurs. Widely used in several domains, including healthcare and churn prediction, a central challenge in its applicability stems from the time of the event being partially observed or right-censoring. Tabular Foundation Models (TFM) have attracted significant interest in recent years due to their ability to perform prediction tasks in a single forward pass, requiring no dataset-specific parameter fitting. Despite their success, their application to prediction tasks on time-to-event data remains difficult due to right censoring. In this work, we present a training-free method to survival regression by leveraging TFMs to both predict the time of the event and iteratively impute right-censored data. Our method uses a TFM to construct an Accelerated Failure Time (AFT) model requiring no training beyond fitting a single scalar parameter. Subsequently, by building on the Buckley-James estimator, we introduce a non-parametric in-context estimator for right-censored data. Our experiments on standard survival analysis benchmarks show that our method is competitive with several parametric and semi-parametric survival regression models that require training, including Cox regression and parametric AFT models.

Mariana Vargas Vieyra

Rhizome Labs, Paris, France

Survival Analysis
AskPDF