acceptodds
Under review as a conference paper at ICLR 2027

Censoring as an Outcome: One Diffusion for Survival Prediction and Cohort Synthesis

Abstract

Survival models estimate when the event happens and treat when the patient left as a nuisance. That discard is why none of them yields a synthetic cohort a statistician can analyse: such a cohort needs the dropout mechanism too, so the target is the law of both latent times. Censoring at random factorises that law and makes it symmetric, an event record right-censoring the censoring time exactly as a censored record right-censors the event time. The censoring channel is then a fitted distribution rather than a weight, and follow-up becomes a quantity one can set instead of a property of the dataset: the model resamples a cohort under a dropout regime it never saw, which is what a power calculation or a planned extension asks for. A model of the observed pair has one lever, truncation, and truncation only shortens. We fit the law with a two-channel score-based diffusion on , trained by variational EM whose E-step draws the unobserved channel from a truncated normal in closed form; the objective upper-bounds the censored-channel negative log-likelihood, and the proposal's slack and the chain's consistency are bounded with it. On six real cohorts the same model leads nine baselines on macro integrated Brier score, time-dependent AUC and Brier resolution, four comparisons resolving at the cohort level, and composed with a tabular backbone it is best on every downstream-utility median we score and beats four competing generators in all six cohorts. Rescaling the fitted dropout curve leaves the event curve inside the estimator's own noise, where truncation matched to the same censoring rate moves it ten times further. Dropping the channel costs IBS, concordance, Brier resolution and reliability in five of six cohorts.

Primary area:
generative models
Keywords:
Generative models, Diffusion Models, Time-to-Event Analysis, Survival Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2vjLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes FATUM, a conditional survival model that learns two times per patient: when the patient would leave the study and when the event (e.g. death) would occur. To achieve this they use a diffusion model with a geometric-Brownian-motion noise schedule so that the forward process is Gaussian in log-time. Since only the earlier time is observed, they train the model with an EM procedure which imputes the missing data. The paper has two goals: to predict survival probabilities and, together with TabDiff as a generative backbone for the patient covariates, to synthesize patient cohorts. Evaluation takes place on six survival datasets against multiple survival models and data generators.

Critical strengths
  • Modelling censoring as a distribution. By treating the censoring time as a latent outcome to be sampled (rather than as a weight) the authors allow the resampling of a cohort under an altered dropout regime. This could potentially be useful for simulation and study planning.
  • Clean technical construction. I especially like the design choice of running the diffusion in log-time so that the censoring constraint becomes an exact truncated-normal draw. This makes the imputation step simple and exact, which gives the method a solid foundation.
Critical weaknesses

1. The framing around survival models is misleading and unsupported, and the paper's goals are unclear.

  • The abstract claims survival models "treat when the patient left as a nuisance. That discard is why none of them yields a synthetic cohort a statistician can analyse." Lines 39–40 state that modelling S(t∣x)S(t \mid x) "is enough to rank patients and not enough to write down a cohort."
  • This is a category mismatch: discriminative survival models are not designed to generate cohorts. The relevant comparison class (SurvivalGAN, SurvDiff, Ashhad & Henao), which already generates censoring patterns, appears only in related work.
  • FATUM itself does not generate cohorts either but requires TabDiff as a backbone for the covariates. Any Table 1 baseline that outputs a survival curve could be composed with TabDiff the same way: one model fit to (Tobs,E)(T^{obs}, E) for events and one fit to (Tobs,1−E)(T^{obs}, 1-E) for censoring. No experiments show that survival models fail at this.
  • The paper pursues two legitimate goals (prediction and cohort generation) but states neither plainly or with a concrete motivation. The abstract relies on undefined terms ("censoring channel", "one lever, truncation") and lists results before establishing what the method is for.

2. Reporting and evaluation do not fully support the claims at the level presented.

  • No standard deviations or per-cohort results. Tables 1 and 2 show only summary values. Per-cohort results exist only for IBS (Table 7), with a standard deviation for FATUM only. No baseline has a standard deviation anywhere in the paper.
  • Median in Table 2. The median is justified by FATUM's own two outlier cohorts (W1W_1 of 1.30 and 0.96 vs. 0.09–0.19 elsewhere). The median hides this failure on a third of the cohorts. Table 1 uses means, so the two tables are not consistent.
  • Mixing scales. Table 1 averages IBS over cohorts where typical values range from 0.06 to 0.20, so a few cohorts dominate the average. The authors' own Appendix E shows the significant results change when the scale is normalised.
  • Generation is evaluated on times only. Table 2 measures fidelity only for the time distribution. There is no covariate fidelity, no check of the relationship between covariates and times, no synthetic censoring rate, and no privacy metric, even though the ethics statement raises privacy.

I would be glad to raise my score if the authors clarify the framing and goals, add the TabDiff + standard survival model baseline, and report per-cohort results with standard deviations (mean and median), together with covariate and joint fidelity metrics.

Questions for authors
  1. How does TabDiff composed with standard survival models (e.g., a parametric AFT and DeepSurv or RSF, each fit to both the event and the censoring distribution) compare to TabDiff + FATUM?
  2. Which cohorts have W1W_1 of 1.30 and 0.96, and what causes the failure?
  3. Do the per-cohort results (mean ± SD for all methods and metrics in Tables 1 and 2) support the summary-level claims?
  4. Why are C-index and RMST gap omitted from Table 2 when they are computed for Fig. 8, and why were TVAE and CTGAN not run with clipped or log-transformed times?
Reviewer_8c68OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FATUM, a diffusion model that learns both event times and censoring times under conditionally independent censoring. It uses an iterative procedure to estimate unobserved times and supports both survival prediction and synthetic cohort generation. The fitted censoring distribution can also be changed to simulate different follow-up conditions

Critical strengths

1- The paper addresses a useful problem. Combining survival prediction with the generation of censored datasets could support research on study design and follow-up strategies. Modelling censoring explicitly gives the generator more control over how subjects leave a study.

2- The experimental evaluation is reasonably broad. It includes six real cohorts, nine predictive baselines, several generative baselines, and controlled simulations. FATUM reports promising average prediction scores and downstream results from models trained on its synthetic data.

3- The authors examine limitations rather than reporting only favourable results. They analyse the approximation used for imputation, include component-removal experiments, and distinguish repeated training seeds from independent datasets in their statistical comparisons.

Critical weaknesses

1- The need for a shared diffusion model is not fully established. Since the method assumes that event and censoring times are independent given the covariates, a natural comparison would fit a separate model for each time and then combine their samples. This comparison would help distinguish the benefit of modelling censoring from the benefit of the proposed diffusion architecture. It is missing from the listed baselines.

2- The generative comparison uses unequal tuning. FATUM’s conditional model receives dataset-specific hyperparameter tuning, whereas the competing generators retain published configurations. The authors explain this choice, but validation-only tuning could also be applied to the baselines. A matched tuning budget would make the claimed advantage more convincing.

3- The benefit of the censoring channel remains uncertain on real data. Removing it worsens prediction error in five of six cohorts, but the differences do not reach significance at the cohort level. The reduced model also inherits settings selected for the full model. An independently tuned event-only version would better establish the practical benefit of adding the censoring channel.

4- The follow-up claims are stronger than the current validation. Changing generated censoring times while leaving generated event times unchanged demonstrates control within the fitted model. It does not, by itself, establish accurate predictions of outcomes that would be observed under longer real follow-up. This matters because the method also performs worse than TabDiff on the reported Wasserstein measure of time-distribution fidelity in all six cohorts.

Questions for authors

1- How does FATUM compare with two separately trained event-time and censoring-time models under the same tuning budget?

2- Would the main results remain stable across different train–test splits, rather than repeated training seeds on one fixed split?

3- Can the follow-up-rescaling experiment be validated against known latent event times or real extended follow-up, rather than mainly checking consistency within the generated data?

Reviewer_hm7QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FATUM, a conditional score-based diffusion model that learns both latent event time �0��0� and censoring time �1��1�. Under the assumption �2��2�, the authors exploit the symmetry of right censoring: an observed event time supplies a lower bound on censoring time, while an observed censoring time supplies a lower bound on event time. The model operates in log-time space and is trained through an approximate variational EM procedure. Missing times are imputed using repeated noise–denoise steps with truncated-Gaussian proposals. The paper analyzes proposal error and finite-chain approximation, including a variance-corrected variant.

The same fitted model supports individual survival prediction and, when combined with TabDiff for covariate generation, synthetic cohort construction. Evaluation includes known-latent simulations and six public survival cohorts, with nine predictive comparators and several generative baselines. FATUM reports the best macro integrated Brier score and time-dependent AUC, and favorable downstream utility from synthetic training data. However, superiority is not established against every comparator, and normalized Wasserstein distance is worse than TabDiff. A further experiment rescales generated censoring times to simulate different follow-up regimes.

Critical strengths
  1. Explicitly fitting both event-time and censoring-time distributions gives the model a concrete role beyond producing a prognostic score. The proposed representation permits generation of complete latent pairs and their corresponding observable records, while retaining an event-survival prediction interface. The log-space construction also makes the censoring constraint straightforward to enforce through truncated-normal sampling. This is a coherent application of diffusion modeling to partially observed outcomes, rather than treating time and event status as arbitrary tabular columns.

This matters because the paper evaluates both uses of the model: prediction on real patients and training downstream survival models on generated cohorts. It therefore asks whether the generated data retain useful conditional relationships, not only whether their marginal distributions appear similar.

  1. Substantial empirical evaluation with explicit reporting of limitations. The experiments use shared splits, multiple training seeds, known-latent simulations, and direct ablations of the censoring channel and variance correction. Table 1 on page 6 reports macro IBS of 0.1437 and time-dependent AUC of 0.7365, while the cohort-level analysis distinguishes these favorable rankings from statistically resolved comparisons. The authors correctly avoid treating all 60 cohort–seed combinations as independent datasets.

The manuscript also retains unfavorable results: the censoring-channel ablation does not establish a cohort-level performance advantage, the variance correction has limited downstream benefit, and time-distribution fidelity is worse than the tabular backbone. This transparency makes the empirical contribution more credible and helps identify what remains attributable to the complete model rather than a specific component.

Critical weaknesses
  1. The proposal-error derivation appears to conflate the ordinary denoising posterior with the censoring-truncated posterior. Let �3��3�. With an exact marginal diffusion score, the Tweedie estimate in Equation 3 corresponds to $
z^=E[z0∣zs,x],\widehat z=\mathbb{E}[z_0\mid z_s,x],

$ whereas the paper defines $

qs=pθ(z0∣zs,x,z0≥a).q_s=p_\theta(z_0\mid z_s,x,z_0\geq a).

$ These are different conditioning statements. Nevertheless, Appendix B.2, Step (b), identifies the difference between the network estimate and the truncated-posterior mean as �4��4� “by Tweedie.” That equality requires clarification of which score defines �5��5�; it does not follow from the marginal score specified in Section 3.1.

A Gaussian-prior sanity check makes the distinction consequential. If the untruncated denoising posterior is �6��6�, the exact Tweedie estimate is �7��7�, and the variance-corrected proposal is $

N(μs,Σs)∣z0≥a=qs.\mathcal N(\mu_s,\Sigma_s)\big|_{z_0\geq a}=q_s.

$

Its KL divergence from �8��8� is therefore exactly zero, including near the truncation boundary. The truncated distribution’s mean shift should not itself be interpreted as a displacement of the Gaussian center used to construct that distribution. This conflicts with the derivation and interpretation of the boundary-dependent location term in Remark 2.

An additional nonnegative term could still yield a loose upper bound, so this observation does not disprove every displayed inequality or invalidate the implementation. It does, however, leave the advertised location-error mechanism and associated quantitative interpretation insufficiently justified. Because theoretical characterization of the approximation is a principal contribution, this needs to be resolved.

  1. The follow-up experiment does not establish the claimed invariance or recovery of outcomes beyond observed follow-up. Section 5.2, P3, argues that leaving the event channel unchanged means its Kaplan–Meier estimate must not move when censoring is rescaled. However, the paper assumes only conditional independence, �9��9�. If both times depend on �10��10�, they can remain dependent marginally, and a pooled, unadjusted Kaplan–Meier curve need not recover the marginal event distribution or remain invariant to changing the censoring regime. Keeping the generated latent event distribution fixed and keeping its Kaplan–Meier estimate fixed are different claims.

The comparison against administrative truncation also needs a common evaluation horizon. Figure 3 on page 8 explicitly shows different endpoints of event-supported estimation followed by plateaus. A large discrepancy evaluated beyond one arm’s supported range can reflect this difference in support rather than superior recovery of the underlying event law. The paper should specify whether its reported tenfold difference persists when comparisons are restricted to shared support.

Finally, conditional independence alone does not identify arbitrary event-time tails beyond the censoring support. For example, with administrative censoring at a fixed time �11��11�, different event distributions can agree through �12��12� and produce identical observations while differing afterward. Generating longer follow-up is therefore a model-based extrapolation unless additional support or tail assumptions are supplied. This matters directly to the proposed study-extension and power-calculation applications: generating a cohort under a changed censoring rule is not, by itself, validation of that cohort’s unobserved outcomes.

Questions for authors
  1. Clarification of the posterior and score used in the theory. In Appendix B.2, is �0��0� the score associated with the ordinary conditional marginal �1��1�, or a score additionally conditioned on �2��2�? How does the equality in Step (b) follow for the network defined in Equation 3? Please reconcile the location term with the Gaussian-prior case, where the exact-variance, exact-score proposal equals the truncated posterior.

  2. Are the Kaplan–Meier curves in P3 adjusted for covariates? Over precisely which time range is the rescaling-versus-truncation discrepancy calculated? Does the principal result persist using a covariate-adjusted marginal estimator on common support, or in a known-latent simulation where recovery under altered follow-up can be checked directly?

  3. Can FATUM be compared with the same TabDiff covariate sampler composed with two separately fitted conditional survival models, one for �3��3� and one for �4��4�, using reversed event indicators for the latter and matched tuning budgets? Such a baseline would distinguish the benefit of diffusion/EM from the benefit of explicitly fitting both distributions.

The related-work positioning should also address MENSA, which already models event and censoring distributions as jointly optimized outcomes. This is not a claim that MENSA evaluates the same cohort-synthesis application, but its explicit modeling of both distributions is relevant to the introduction’s broad exclusivity claims.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.