Adaptation Interfaces for In-Context Tabular Foundation Models in Time-to-Event Prediction
Minh-Khoi Pham, Luca Cotugno, Dan Cernei, Alina Sirbu, Stefano Masi, Giuseppe Prencipe, Alessandro Pingitore, Patrizia Landi, et al.
cs.LG · cs.AI
Quick verdict
Compares three TabFM-to-survival adaptation interfaces on 74+4 datasets: careful, statistically backed, but competing-risk conclusions rest on just 4 datasets.
Summary
This benchmark study asks how generic tabular foundation models (TabPFN, TabDPT, TabICL) should be adapted for censored time-to-event prediction. Building on the authors' prior MTLR-head work, three transfer interfaces are compared on 74 public single-risk datasets plus 4 competing-risk datasets: zero-shot horizon-wise classification reformulation, censoring-aware classification fine-tuning on temporally expanded examples, and survival-head adaptation (CoxPH, DeepHit, MTLR) over frozen backbones with context-resampled training. Evaluation uses the Antolini time-dependent concordance index and Integrated Brier Score under 5-fold cross-validation, with paired Wilcoxon tests, Benjamini-Hochberg FDR correction, and Friedman tests across backbones. Zero-shot inference is competitive on small datasets while supervised adaptation wins as scale grows; Cox is the most consistently strong head, especially for IBS on large datasets; DeepHit favors discrimination over probabilistic accuracy; and cause-specific MTLR ranks highest among TabFM heads in the four-dataset competing-risk arm. Notable limitations: competing-risk conclusions rest on only four datasets, the large-data regime contains nine datasets, and TabFM variants use fixed hyperparameters while baselines receive 20-trial Optuna tuning, complicating fairness. Code is released. Overall a careful, well-hedged empirical contribution whose headline orderings are framed as regime-dependent rather than universal.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Zero-shot in-context reformulation is effective on smaller single-risk datasets, while supervised adaptation becomes increasingly advantageous as dataset size grows.
- Among survival heads on frozen TabFM backbones, CoxPH is the most reliably strong interface, especially for Integrated Brier Score on larger datasets.
- DeepHit adaptation is relatively stronger for time-dependent concordance than for IBS, consistent with the ranking component of its loss.
- In the four-dataset competing-risk analysis, cause-specific MTLR ranks highest among the TabFM survival heads.
- Classification fine-tuning narrows the gap to zero-shot as data grows but remains weaker than survival heads for probabilistic prediction.
Strengths
Broad, heterogeneous benchmark with regime stratification
74 SurvSet-derived single-risk datasets plus 4 competing-risk sets (SUPPORT2-CR, FRAMINGHAM, PBC2, SYNTHETIC), stratified into 40 small / 25 medium / 9 large (Sec. 4.1, Fig. 2), which is what makes the regime-dependence conclusion interpretable rather than anecdotal.
Real statistical testing, not just mean ranks
Appendix D reports dataset-level paired one-sided Wilcoxon tests with Benjamini-Hochberg FDR (q<=0.05), win scores with a 0.001 tie band, and Friedman omnibus tests across backbones (DeepHit q=0.030; Cox q=0.053; MTLR q=0.149), with full pairwise win matrices (Fig. 4).
Leakage-aware, principled method engineering
Matched twins isolate the backbone effect
Unusually honest hedging plus useful ablations
Weaknesses
Asymmetric tuning budget across model families
Competing-risk headline rests on four datasets
Large-regime claims stand on nine datasets
Context labels discard event time
Survival-pretrained FMs positioned but not compared
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper reframes tabular-foundation-model (TabFM) survival transfer as an *adaptation-interface* problem and benchmarks three routes on frozen backbones (TabPFN, TabDPT, TabICL): (i) zero-shot horizon-wise censoring-aware classification (Secs. 3.2-3.3, following Kim et al. 2026), (ii) classification fine-tuning (-CE) on subject-horizon expanded examples with validity masks and 0.5 negative subsampling, and (iii) survival-head adaptation (-Cox/-DH/-MTLR) trained under context-resampled support sets of up to 512 examples with R=5 inference ensembles (Sec. 3.2, Algorithm 1). Headline claims: zero-shot is competitive on small single-risk sets; supervised adaptation wins as scale grows; Cox is the most reliably strong head, especially for IBS on large sets; DeepHit favors discrimination (Ctd) over probabilistic accuracy; cause-specific MTLR tops the four-dataset competing-risk arm. Claims are unusually well-bounded by regime and metric, and Sec. 6 explicitly disclaims universal orderings.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are adaptation interfaces for tabular foundation models in survival analysis?
They are the different ways to connect a pretrained tabular model (like TabPFN, TabDPT, or TabICL) to censored time-to-event prediction: using it zero-shot through horizon-wise binary classification, fine-tuning a classification head on time-expanded examples, or attaching a dedicated survival head (Cox, DeepHit, MTLR) to the frozen backbone. This paper benchmarks all three on 74 single-risk and 4 competing-risk datasets.
Which adaptation strategy works best for survival prediction with TabPFN-style models?
It depends on dataset size. Zero-shot inference is competitive on small datasets; as data grows, supervised adaptation wins, and a Cox head gives the most consistently strong results — especially for the Integrated Brier Score on large datasets. DeepHit is relatively better at ranking (concordance) than at calibrated probabilities.
Is the code and data for this paper available?
Yes. The paper links implementation scripts at github.com/kaylode/survival-fm and uses public data: SurvSet for the 74 single-risk datasets, plus SUPPORT2-CR, FRAMINGHAM, PBC2, and SYNTHETIC for the competing-risk experiments.
Can the competing-risk conclusions be trusted?
Only tentatively. The competing-risk comparison covers just four datasets, and the authors themselves state this is too few to establish a stable ordering — even though cause-specific MTLR ranked highest among the tested heads on both concordance and Brier-based metrics.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with [...sampling gap...] regions: portions of Sec. 3.4 (survival-head loss definitions), parts of the Discussion/limitations, Appendix A/B, and chunks of Tables 5-6 were skipped; claims resting solely on those regions were avoided.
- Per-dataset Table 5 rows are only partially visible (alphabetical subset), so dataset-level spot checks were limited to the entries actually shown (e.g., GBSG2, FLCHAIN, COLON).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.