acceptodds
Under review as a conference paper at ICLR 2027

Reward-Guided Survival Data Generation

Abstract

The sharing of tabular survival data is restricted by privacy regulations. While synthetic data generation offers a promising alternative, existing methods frequently fail to simultaneously preserve long-term survival dynamics and patient privacy. We introduce SurvTabGen, a novel generative framework tailored specifically for survival data synthesis. Unlike prior multi-stage approaches, SurvTabGen jointly models covariates, event indicators, and survival time within a unified structurally aware generator. We first extract causal relationships among covariates via causal discovery. The generation then follows a dependency-aware sequence: event outcomes are conditioned on covariates, and survival times are conditioned on both covariates and events. Crucially, the generator is trained with a reinforcement learning-based survival-aware objective. By using the Integrated Absolute Difference (IAD) between real and synthetic survival curves as a reward signal, the model directly optimizes for macroscopic temporal fidelity. Extensive evaluations against 10 SOTA baselines across seven diverse real-world datasets demonstrate the superiority of SurvTabGen. Our results demonstrate that SurvTabGen provides a highly effective solution for generating high-fidelity, privacy-preserving synthetic cohorts that excel in survival preservation and data quality.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Survival Data, Synthetic Data, Causal, Kaplan-Meier, Generative Models, GANs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dwVAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SurvTabGen, a structured generative model for synthetic survival data. It combines adversarial training with a policy-gradient reward based on the discrepancy between real and synthetic Kaplan–Meier curves. Experiments show improved population-level survival-curve agreement and competitive downstream prediction.

Critical strengths

The method directly incorporates survival-curve agreement into generator training. The experiments and ablation study demonstrate that this objective improves the corresponding population-level metric.

Critical weaknesses
  1. The reward is insufficiently justified for the intended use of synthetic data. The Wasserstein identity explains the marginal survival-curve objective, but does not establish that optimizing it preserves conditional survival distributions. Different subpopulations can have incorrect or reversed risk relationships while the overall survival curve remains correct. The limited subgroup evaluation does not resolve this concern.

  2. The motivation for generating an entire dataset is not clearly connected to the main objective. If the intended use is only to recover one population survival curve, sharing that curve would be a simpler alternative. Synthetic data becomes useful when it supports additional analyses, such as subgroup risks or covariate–outcome relationships. The paper needs stronger evidence that its reward improves these uses beyond improving the curve it directly optimizes.

Questions for authors
  1. Under what assumptions does reducing the marginal survival-curve discrepancy also control errors in subgroup or conditional survival distributions? If there is no such implication, what prevents the reward from favoring aggregate agreement at the expense of subgroup accuracy?

  2. Which intended analyses require synthetic records rather than a released survival curve? For those analyses, does adding the survival reward improve accuracy compared with the same generator without that reward?

4 comments
AuthorsOpenReview ↗

Response to Q1: Marginal Survival Curve vs. Subgroup-Conditional Distributions

We sincerely thank Reviewer dwVA for the thoughtful and technically insightful feedback. We respond to your two questions below.

Q1: Marginal Survival Curve vs. Subgroup-Conditional Distributions

First, while preserving survival patterns across subpopulations is important, subgroup definitions are application-dependent and often require domain expert input to ensure clinical or practical relevance. Because meaningful subgroups vary widely across the seven diverse clinical and industrial datasets evaluated in this work, a unified subgroup evaluation protocol does not exist. Our main objective is therefore to preserve the global survival curve on the whole population, aligning with the objective of other state-of-the-art methods such as SurvivalGAN, Ashhad, and SurvDiff.

Second, we agree that minimizing the marginal survival-curve discrepancy (IAD) does not, by itself, theoretically guarantee preservation of subgroup-conditional survival distributions. However, the survival reward is not the sole component of SurvTabGen's training objective. Our generator jointly optimizes LG=Ladv+λLsurv\mathcal{L}_{G}=\mathcal{L}_{\text{adv}}+\lambda\mathcal{L}_{\text{surv}}. The adversarial objective operates on the complete generated records (X,E,T)(X,E,T), encouraging the synthetic joint distribution to match the real joint distribution. In addition, our dependency-aware architecture generates event indicators conditioned on covariates and survival times conditioned on both covariates and event indicators. Thus, the adversarial objective and conditional generation architecture complement the marginal survival reward by encouraging preservation of covariate-outcome dependencies.

In particular, at the population level, exact matching of the joint distribution would imply matching conditional distributions almost everywhere on the support of the covariates. Although finite-sample adversarial training does not guarantee exact matching or provide a formal bound on subgroup errors, it discourages solutions that reproduce the global survival curve while substantially distorting covariate-outcome relationships.

Finally, we also provide empirical subgroup-level evidence in Appendix E.2 (Table 5). On the Whas500 dataset, SurvTabGen achieves the lowest KM Divergence for the overall population (0.035), as well as the Elderly (0.076) and Very Elderly (0.079) cohorts. SurvivalGAN performs best for the Young cohort (0.063), while SurvTabGen obtains 0.131. While we acknowledge that subgroup preservation is not uniformly superior across all cohorts, these results provide evidence that optimizing the global survival objective can be compatible with preserving subgroup-specific survival patterns.

AuthorsOpenReview ↗

Response to Q2: Justification for Synthetic Data and the Impact of the Reward on Downstream Utility

Q2: Justification for Synthetic Data and the Impact of the Reward on Downstream Utility

We agree that releasing a population-level Kaplan-Meier curve would be sufficient if the sole objective were to communicate aggregate survival behavior. However, the purpose of synthetic record generation is broader: to provide reusable, individual-level records that retain covariates and their relationships with event outcomes and observed times. Such records enable analyses that cannot be performed using an aggregate survival curve alone, including covariate-based risk stratification, subgroup comparisons, and training downstream survival prediction models.

We demonstrate this utility in Appendix E.7 (Table 8), where three survival models (CoxPH, SurvivalXGBoost, and Random Survival Forest) are trained on synthetic records and evaluated on held-out real data. SurvTabGen achieves an average C-index of 0.703, close to SurvDiff (0.711), and an average Brier Score of 0.134, close to the best-performing baseline, TapTap (0.126). This performance is highly competitive, falling just slightly behind the best methods by a narrow margin of 0.008 on both metrics. More importantly, SurvTabGen outperforms or remains competitive with models trained directly on the real data. These results demonstrate that SurvTabGen generates records that support downstream survival prediction, rather than merely reproducing an aggregate survival curve.

To answer the reviewer's question whether the survival reward contributes to this downstream utility, we conducted an additional ablation comparing the same generator with λ=0\lambda=0 (no reward) and λ=0.1\lambda=0.1 (with reward), using the downstream evaluation protocol in Appendix E.7. The results are reported below.

The survival reward improves the average C-index from 0.684 to 0.703, with improvements on six of the seven datasets and no change on Flchain. The average Brier Score also decreases from 0.190 to 0.134, although the improvement is not uniform across datasets. These results indicate that the survival reward contributes to downstream predictive performance.

C-index (↑\uparrow, higher is better)

DatasetSurvTabGen (λ=0\lambda=0)SurvTabGen (λ=0.1\lambda=0.1)
AIDS0.697 (0.05)0.707 (0.05)
Cgvhd0.500 (0.00)0.551 (0.09)
Flchain0.929 (0.00)0.929 (0.00)
Gbsg20.638 (0.02)0.656 (0.02)
Veterans0.605 (0.09)0.641 (0.03)
Whas5000.721 (0.03)0.728 (0.01)
AM0.700 (0.04)0.711 (0.05)
Average0.6840.703

Brier Score (↓\downarrow, lower is better)

DatasetSurvTabGen (λ=0\lambda=0)SurvTabGen (λ=0.1\lambda=0.1)
AIDS0.068 (0.01)0.061 (0.00)
Cgvhd0.515 (0.20)0.129 (0.02)
Flchain0.050 (0.00)0.050 (0.00)
Gbsg20.269 (0.01)0.272 (0.01)
Veterans0.162 (0.04)0.144 (0.03)
Whas5000.194 (0.01)0.199 (0.01)
AM0.077 (0.01)0.083 (0.01)
Average0.1900.134

We hope these new empirical results address your concerns, and we kindly ask that you reconsider your evaluation.

Dear Reviewer dwVA,

As the Author/Reviewer discussion period concludes today, Nov. 18, we want to follow up to see if our rebuttal has successfully resolved your concerns.

We would be happy to answer any remaining questions.

If our responses have addressed your points, we would be grateful if you considered reflecting this in your overall evaluation.

Thank you again for your time and constructive feedback.

Best regards,

The Authors

Thank you for the careful response and the additional ablation, which I appreciate. I am afraid my main concern is not yet fully resolved. Since the overall survival curve can be obtained directly from the real data, I am still unsure why the generator should be rewarded for matching it, and I could not find clear evidence in the response that this reward helps the uses that need synthetic records, such as subgroup and covariate-outcome analysis, rather than trading them off. I would therefore like to keep my score for now.

Reviewer_TkzfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce SurvTabGen, which generates survival data in a unified structure that models the covariates, event times, and event indicators simultaneously. Their approach first builds a causal graph from data and then produces covariates based on it, which are then used to condition events, and subsequently event times. The generator is trained through reinforcement learning (RL) to match the population-based event distribution. Results on 7 datasets and 10 baselines are used to demonstrate the capabilities of the proposed model.

Critical strengths

The use of causal discovery to generate covariates is intriguing.

Critical weaknesses

The authors state that the reverse-conditional distribution used by other approaches is structurally more difficult to optimize than the forward conditional, however, do not acknowledge that in their approach their perform causal estimation which is arguably a much harder problem than modeling the conditional distributions required for survival analysis.

The authors note the objective to match survival distributions as a contribution, which based on the specific method the use based on RL may be new, however, the matching of such distributions has been previously done in the context of calibration for survival analysis models and without the need for RL objectives (P. Chapfuwa et al., Calibration and Uncertainty in Neural Time-to-Event Modeling in IEEE TNNLS, 2023).

The authors claim privacy as a driver for the need for synthetic survival data and argue that existing methods may not be privacy preserving, however, they do not provide guarantees about how the proposed method is in any way privacy preserving other than the results in Table 2, which though welcome are not definitive.

The experiments are underwhelming: i) using KM as the primary metric could be perceived as “metric hacking”, ii) results in Figure 3 only show two baselines and iii) results in Figure 4 and Table 2 do not present error bars.

Questions for authors

Reconsider the statement about the 10 SOTA baselines provided that as the authors argue, there are only three models specifically designed to generate survival data.

It is clear that imposing an inductive bias by making X be the parents of E and (X,E) be the parents of T is reasonable by temporality, however, it is not clear why all covariates in X are considered parents, which is likely to introduce undesirable bias.

Why are not other methods included in Figure 3 and why not showing the other datasets, especially Ashhad which ranks second in Table 1.

Without error bars, results in Figure 4 and Table 2 are difficult to interpret.

Can the authors argue for the reason behind the loss in KM divergence in Figure 5 when only removing the causal structure?

3 comments
AuthorsOpenReview ↗

Responses to Weaknesses 1-3 and Question 5

We sincerely thank Reviewer Tkzf for the thoughtful and technically insightful feedback. We address the concerns and questions below.

W1 & Q5: Causal Discovery vs. Reverse-Conditional Generation

We clarify that causal discovery in SurvTabGen is applied exclusively to the covariates XX, with the aim of learning their dependency structure for conditional generation. We do not use causal discovery to estimate the relationships between XX, EE, and TT. Instead, the generation order X→E→TX\rightarrow E\rightarrow T is imposed as a survival-oriented inductive bias.

We acknowledge that estimating a covariate dependency graph introduces its own challenges, including the possibility of graph misspecification. Our motivation is therefore not to claim that causal discovery is inherently easier than conditional distribution estimation, but to use a covariate-first generation architecture that preserves learned dependencies among baseline features. This avoids the outcome-conditioned covariate generation used by prior approaches (e.g., P(X∣E)P(X\mid E) in Ashhad), although such reverse-conditioned approaches are also valid generative factorizations.

To assess the sensitivity to graph estimation, we provide two analyses. First, Appendix E.4 compares alternative causal discovery methods GES, FCI, and BOSS, showing that SurvTabGen maintains competitive survival fidelity across different graph-learning methods. Second, Appendix E.6 evaluates progressively perturbed graph structures. Although severe graph misspecification degrades KM Divergence, SurvTabGen remains reasonably robust even when graph accuracy decreases to 20%. These results do not establish that the estimated graph is the true causal structure, but demonstrate that the framework is not dependent on a perfectly recovered graph.

To answer Q5 (regarding the loss in KM divergence when removing the causal structure): In the "No causal structure" ablation, we remove all learned dependencies among the covariates by treating each covariate as an independent node with an empty parent set. Consequently, each covariate is generated independently from its own noise input. Meanwhile, the event and time generators remain conditioned on the generated covariates. Without the learned covariate dependencies, correlations among baseline features are not preserved, producing unrealistic combinations of covariate values. These combinations are then passed to the conditional event and time generators, changing the resulting event-time relationships. This explains why removing the covariate dependency structure degrades the average KM Divergence from 0.026 to 0.037.

W2: Prior Work on Survival-Function Matching

We thank the reviewer for highlighting Chapfuwa et al. (2023). We agree that this is relevant prior work and will include it in the Related Works section.

We clarify that the novelty of SurvTabGen does not lie in survival-function matching alone. Chapfuwa et al. use survival-function matching to improve the calibration of conditional time-to-event predictions. In contrast, SurvTabGen addresses the generation of complete synthetic survival records, including covariates, event indicators, and observed survival times. Specifically, we integrate a non-differentiable, population-level Kaplan-Meier based objective into a structured joint-data generator and optimize this reward through policy gradients (REINFORCE).

W3: Privacy Guarantees

Our focus is on empirical privacy evaluation, aligning with the standard evaluation protocol established by recent survival data generators like Ashhad. While Ashhad evaluates privacy protection empirically using only Distance to Closest Record (DCR), other methods such as SurvivalGAN and SurvDiff do not evaluate empirical privacy preservation in their main results. In contrast, SurvTabGen evaluates empirical privacy using both DCR and Re-identification Risk. We also note that while SurvDiff discusses Differential Privacy as an extension, its primary evaluation follows a non-DP setting.

Following your suggestion, we have updated Table 2 in the revised paper to include the standard deviation for both privacy metrics.

AuthorsOpenReview ↗

Responses to Weakness 4 and Questions 1-4

W4: KM Divergence and the Concern of Metric Hacking

We clarify that while KM Divergence serves as our primary survival metric, our empirical evaluation is comprehensive and extends far beyond a single metric. We evaluate SurvTabGen using 10 metrics across 5 dimensions: Survival Fidelity (KM Divergence, Optimism, Short-sightedness), Downstream Utility (C-index, Brier Score), Data Quality (Jensen-Shannon and Wasserstein distances), Privacy (Distance to Closest Record, Re-identification Risk), and Computational Efficiency. We also provide qualitative visual assessments via Kaplan-Meier curves and t-SNE embeddings.

Furthermore, the scope of our evaluation far exceeds recent baselines. For instance, while SurvDiff (ICML 2026) was primarily evaluated on Downstream Utility and Data Quality across just 3 clinical datasets, SurvTabGen was tested across 7 diverse datasets spanning both healthcare and engineering domains.

W4, Q3 & Q4: Additional Survival Curves, Datasets, and Error Bars

Thank you for these suggestions. We have substantially expanded the visual and statistical presentation in the revised paper.

First, Figure 3 previously compared only SurvivalGAN and SurvTabGen (as we consider SurvivalGAN our primary GAN-based competitor). Figure 3 now includes Kaplan-Meier curves for all 4 survival-specific methods (SurvivalGAN, Ashhad, SurvDiff, and SurvTabGen) across all 7 datasets. As shown in the updated figure, SurvTabGen closely preserves the real survival trajectories on most datasets. In contrast, SurvDiff frequently suffers from over-pessimism (visible in 4/7 datasets), while Ashhad suffers from short-sightedness on 3 datasets (Cgvhd, Veterans, and AM).

Second, Figure 4 now includes error bars for the Jensen-Shannon and Wasserstein distances across all methods. We have also updated Table 2 to include standard deviation for DCR and Re-identification Risk.

Q1: Clarification of the 10 Baselines

We agree that it is necessary to distinguish general-purpose tabular generators from survival-aware generators. Therefore, we will reword "10 SOTA baselines" to "10 baselines, including three methods specifically designed for survival data generation".

Q2: Inductive Bias of PAE=XPA_{E}=X and PAT=X∪{E}PA_{T}=X\cup\{E\}

We thank the reviewer for raising this important question. We acknowledge that designating all covariates in XX as parents may introduce a dense inductive bias. However, because our objective is to model the true joint distribution P(X,E,T)P(X,E,T), which factorizes as P(X)P(E∣X)P(T∣X,E)P(X)P(E\vert X)P(T\vert X,E) via the chain rule of probability, it is natural to utilize the complete covariate set XX as the conditioning parent set for both EE and TT. Furthermore, this design uses the complete set of observed baseline covariates to avoid prematurely excluding potentially informative features when generating event indicators and survival times.

We hope these revisions and clarifications address your concerns and highlight the robustness of our framework. We kindly ask you to reconsider your overall evaluation.

Dear Reviewer Tkzf,

As the Author/Reviewer discussion period concludes today, Nov. 18, we want to follow up to see if our rebuttal has successfully resolved your concerns.

We would be happy to answer any remaining questions.

If our responses have addressed your points, we would be grateful if you considered reflecting this in your overall evaluation.

Thank you again for your time and constructive feedback.

Best regards,

The Authors

Reviewer_kj5HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to generate synthetic data (SD). Specifically they assume a structure for this problem where there are covariates XX, a binary event EE, and a time TT when that event occurs (and there can be censorship). The goal is to generate samples from P(X,E,T)P(X,E,T).

The motivation for generating SD is that collecting real data is time-consuming and often must be kept private, and generating SD is fast and preserves privacy allowing users to share it. Recent works (CTGAN, TabDiff, and GReaT) for generating tabular SD do not account for sensoring in survival time-series data, exhibit 3 common failure modes: (1) over-optimistism—events occur later than they should (2) over-pessimistism—events occur earlier than they should (3) short-sightedness—model fails to generate long horizons.

Only three models have been proposed for survival data generation: Ashhad, SurvivalGAN, and SurvDiff. The first two sample E,TE,T from the empirical data distribution, which risks not preserving privacy, as the model is direcly sampling real events outcomes and times. Given those empirical samples the model admits sampling from P(X∣E)P(X|E), thus is sampling the covariates given the future outcome. The paper claims this reverse-conditional generation (of past covariates given the future outcome), is "structurally more difficult to optimize" and "violates the natural chronological progression". The last model, SurvDiff, samples P(X,E,T)P(X,E,T) jointly using diffusion, and it's drawbacks are (1) it treats all features as structurally independent (2) it's loss is not a global survival curve loss. These drawbacks lead to it suffering from failure models 1-3 above. To generate SD that aligns with the data generating process (the forward process), this method factorizes P(X,E,T)P(X,E,T) as P(T∣E,X)P(E∣X)P(X)P(T|E,X)P(E|X)P(X), sampling XX before conditionally sampling EE and then TT. Additionally this method learns a causal DAG for the covariates XX. To address the surival curve failure modes 1-3, the authors use Integrated Absolute Difference (IAD) between the model and real Kaplan-Meier curves. This loss is non-differentiable and is fitted using policy gradients.

Contributions:

  1. Framework for generating survival SD via a Structural causal model matching the data generating process.
  2. Integrated Absolute Difference (IAD) loss optimized via policy gradients
  3. Empirical validation: Better survival fidelity
  4. Avoids complex multi-stage model training or reverse-conditional estimation of Ashhad and SurvivalGAN.
  5. This method respects causal feature relationships unlike SurvDiff (assuming that the PC algorithm idenifiability assumptions are satisfied for the covariates)

Learning Algorithm summary:

A discriminator is trained using the same loss as in W-GAN-GP. Generator generates data autoregressively using the PC algorithm learned covariate DAG topological ordering, and subsequently generating E and finally T. An adverserial loss and survival loss are used to train this generator. Adversarial loss equates to fooling the discriminator: just minimizing D(u^)D(\hat u). The survival loss (aka IAD) is a 1-Wasserstein distance between the population survival distribution and generator's survival distribution. Empirically it is computed on KM estimators and using a batch and the integration is computed over a grid T\mathcal T. Since the survival does not admit differentiation of generator weights, the REINFORCE estimator is used which requires being able to evaluate the loss and evaluate the dirivative of the log-liklihood (score) from the generator, which this paper uses a surrogate for. REINFORCE typically just directly multiplies the loss by the score, however this paper subtracts P(B)\mathcal{P(B)} (a loss term evaluating how accurately the model is matching quantiles for TT and the probability of E=1E=1 between real and generated data) from the loss prior to multiplying by the score.

Critical strengths

The KM-Divergence results are relatively good, the model beats all other methods (except ashhad and SurvivalGan) in terms of KM on all taks except AM where it is beat by Survtabgen, PredLLM, Great, and Survdiff. It's losses to Ashhad and SurvivalGan are defensible given significant improvements in DCR metric for privacy preservation, although the Risk metric shows a more modest difference.

The paper is well written, relatively easy to follow, contains a lot of results, and provides a this novel IAD loss term which is an intuitive global loss term.

Critical weaknesses

Evaluation:

Jensen-Shannon and Wasserstein distances are used to evaluate synthetic-data quality, however these metrics are simply averaged across the datasets and I could not find in the manuscript the per-dataset results. Additionally the model is beat by both Great and PredLLM for both of these metrics when we just look at this average. Due to simpsons paradox it is important to look at the raw data and not trust the average. I am curious what the win-rate is of the proposed approach across tasks (what percent of the tasks does it beat all other methods on), and what is the statistically-significant win rate (percent of tasks it wins by a statistically significant amount).

The only other syntheitic-data-quality metric in the paper is in appendix E.7 which evaluates this approach vs other SD approaches in train on synthetic test on real (TSTR) setting. In this evaluation, the proposed approach is never the best for Brier nor for C-Index across all tasks, implying this model may not be learning structure useful for predictive tasks, while other models are.

The current privacy evaluation is challenging to interpret. Below I list the two metrics and the challenges in guaging if the scale of improvement with this approach is meaningful: Reidentification risk: .307 is the reidentificaiton risk, so 30.7% of the time, if I take a real patient, a synthetic sample is closer to it than a real sample. On a particular dataset, with M synthetic samples and N real samples, if the synthetic data follows the exact unerlying distribution and is not memorizing, then we expect a proportion of MM+N−1\frac{M}{M + N -1} to be the fraction of the time that a synthetic patient is the nearest neighbor, and if this value goes up then there is memorization. Could the authors provided this to contextualize the reidentification risk since the scale is uninterpretable. The risks should be provided for each dataset with the baseline described. As the results currently stand it is hard if the reduction in reidentification risk are meaningful. if MM+N−1\frac{M}{M + N -1} is close to 0 then a 2-3% improvment over Asshad and survivalGan (which are sampling real E,T) is not very meaningful, while if it is close to .29 then it is a more meaningful improvement.

DCR: Hard to interpret DCR value on it's own. It is a measure of on average how far a generated synthetic patient is to the closest real patient. Perhaps it would be more meaningful to compare this number computed on training-set patients, to patients in a held-out set (where there is no memorization/privacy risk). Then the scale of the DCR is more meaningful, as we see how this distance get's lower due to memorization alone.

Applicability

Datasets can have multiple time and event columns and subsets of these may be competing and terminal, and some may occur multiple times before censorship, this framework only allows a single time column and a single corresponding event column, and the event can only occur once. Covariates also must be tabular and can not have a time-varying component, thus users are expected to formulate a tabular feature vector of covariates paired with the time & event columns, which limits the type of data that is modeled and generated.

Questions for authors

To raise my score I would need to see consistent improvements in general data quality metrics (JS, Wasserstein distances) where this model is winning most of the time on tasks, if this is not the case the authors may need to significantly modify this approach and resubmit.

  • Can the data-quality evaluation questions in the weakeness section be addressed?
  • Can the privacy evaluation questions in the weakenss section be addressed?
  • For the applicability weakness, could the authors comment on if this can be extended to use cases with multiple times and multiple event types that may be competing (ex. different causes of death), or may not be competing (ex. is medical labs).

There is a bit of lack of clarity for a few things in the current manuscript, I added questions below for them:

  • It is claimed that P(B)\mathcal{P(B)} regularizes the policy gradients but I think there is no ablation nor further intuition clarifying why this particular term is added. It is just matching quantiles for TT and the probability of E=1E=1 between real and generated data. Could the authors clarify why this was chosen, is there a precedent in other literature or was this an empirical finding they discovered?
  • Can the authors clarify if they used the Gumbel-softmax or straight-through softmax. It is unclear how categorical samples are treated, i.e is the generator outputing a continuous risk for each category and then a discrete sample is sampled and provided to generators of it's children, or is the soft risk being provided. Additionally what temperature is used?
  • On line 286, the authors say for the surogate liklihood: "For continuous variables, the generator outputs parameterize a stochastic distribution (e.g., Gaussian)." It is unclear what actual distribution is being used for this surrogate. Is it gaussian, it it fitted using samples stochastically generated by varying ziz_i?.
  • In figure 2, what does an alive event mean? In the Survival Data table in the figure, the second row shows an alive event at time 24. Is the meaning that the patient at time 0 is 42, has a surgery planned, and then at time 24 the patient is alive? Is the surgery at time 24.
  • What is the actual T\mathcal T values used for each dataset. I could not find this in the paper.

Nits:

For figure 6, it would be interesting to see if T-SNE for subgroups where T is in a particular range. Then one can see if structure is better preserved between the population and generated data at different cross-sections/bins of time at risk.

A description of why KM is not differentiable by TiT_i would help this read more easily, something like this perhaps where djd_j and njn_j are explicitly broken up into the the indicator functions of TjT_j could make it more obvious to a reader:

This empirical quantity is not differentiable in terms of TjT_j since KM takes the form S(t)=∏ti≤t(1−∑jI{Tj>ti}Ej∑jI{Tj>tj})S(t) = \prod_{t_i \leq t} (1 - \frac{\sum_j I\{T_j > t_i\} E_j}{\sum_j I\{T_j> t_j\}}), where ii indexes all samples and jj indexes the times in our grid T\mathcal T which are less than tt. This is not differentiable because the indicator function is a step function for TjT_j and has zero gradient everywhere except at time tt. It may also help to clarify that KM is differentiable by EiE_i, but, as the paper mentions, the sampling process is not thus necessitating the gumbel-softmax trick.

I don't think line 268 is supposed to be italicized right? It discusses the proposition but is not part of it's claim.

6 comments
AuthorsOpenReview ↗

Responses to Weaknesses 1-5

We sincerely thank Reviewer kj5H for the thoughtful and technically insightful feedback. We address your main concerns below. For your minor comments, we will incorporate the clarifications into the final version.

W1: Data Quality (JS/WS Distances) vs. Survival Fidelity

Although distributional fidelity is an important dimension of synthetic data evaluation, the primary objective of SurvTabGen differs from that of general-purpose tabular data generators. While methods such as GReaT and Pred-LLM focus on reproducing general tabular distributions, SurvTabGen is designed to generate synthetic survival records that preserve population-level survival dynamics under censoring. Accordingly, our evaluation emphasizes two critical aspects: (1) survival fidelity, measured by KM Divergence, Optimism, and Short-sightedness and (2) downstream utility, measured by C-index and Brier Score.

As you observed, GReaT and Pred-LLM achieve lower average JS and Wasserstein distances. However, these improvements do not translate into better survival-specific performance. For example, SurvTabGen achieves a substantially lower average KM Divergence (0.026) than GReaT (0.055) and Pred-LLM (0.050). It also achieves better average downstream C-index (0.703 vs. 0.611 and 0.665) and Brier Score (0.134 vs. 0.184 and 0.209), respectively.

Because our goal is to generate synthetic survival data rather than generic tabular data, optimizing for a JS or Wasserstein distance "win-rate" is misaligned with our primary objective.

(Note: Following another reviewer's suggestion, we have added error bars to Figure 4 to show the variability of JS and Wasserstein distances across datasets).

W2: Downstream Performance (TSTR)

We acknowledge that SurvTabGen does not achieve the best C-index or Brier Score on every individual dataset. Nevertheless, it achieves competitive average downstream performance, with a C-index of 0.703 and a Brier Score of 0.134. The corresponding best average scores are 0.711 for SurvDiff (C-index) and 0.126 for TapTap (Brier Score), resulting in a negligible absolute differences of 0.008 for both metrics.

More importantly, SurvTabGen achieves this utility with superior computational efficiency. While SurvDiff requires significantly longer training (1,200 epochs, ~26 mins/dataset), SurvTabGen only needs 500 epochs (~3 mins/dataset) (Figure 10).

W3: Re-identification Risk

We appreciate the reviewer for suggesting a reference baseline Rnull=MM+N−1R_{\mathrm{null}}=\frac{M}{M+N-1}, to contextualize the re-identification risk. In our setting (see Problem Formulation), the number of generated synthetic records equals the number of real records (M=NM=N). Therefore, we can derive the baseline threshold: Rnull=N2N−1≈0.5R_{\mathrm{null}}=\frac{N}{2N-1}\approx0.5.

Because this threshold 0.50.5 applies across all datasets, SurvTabGen's average risk score of 0.3070.307 confirms that the model is generating novel synthetic records rather than memorizing the training data. Furthermore, it demonstrates that our 2-3% improvement over SurvivalGAN and Ashhad is meaningful.

W4: DCR Evaluation

We thank the reviewer for suggesting the held-out-real-record DCR reference. However, we compute the DCR score between real training set and synthetic set for two reasons. First, it is a standard privacy evaluation protocol for tabular synthesis (used by Ashhad, GReaT, and Pred-LLM). Second, to compute evaluation metrics (with the exception of the downstream predictive tasks in Appendix E.7), we train generators using the complete real dataset and generate a synthetic dataset of the same size. Therefore, there is no held-out real set available during the standard DCR computation.

However, as an additional reference, we calculated the average DCR between held-out real records and the corresponding training records under our 67/33 downstream evaluation split. The resulting average Test-to-Train DCR is 0.072 (0.07).

(Note: Following another reviewer's suggestion, we have also added the standard deviations for DCR and Re-identification Risk to Table 2).

W5: Applicability to Multiple/Recurrent Events

SurvTabGen currently focuses on right-censored survival data with a single event of interest and static tabular baseline covariates. It covers the majority of clinical survival datasets and aligns with the benchmarking scope of prior survival generators (SurvivalGAN, Ashhad, SurvDiff). Extending the framework to competing risks would require modeling multiple event causes and their corresponding cumulative incidence functions. Recurrent-event and longitudinal settings would require sequential generation of multiple event times or time-varying covariates. These are different statistical modeling tasks, which we consider important extensions for future work.

AuthorsOpenReview ↗

Responses to Questions 1-5

Answers to Specific Questions:

Q1. Auxiliary Penalty P(B)\mathcal{P}(\mathcal{B}): The auxiliary penalty is a simple reward-shaping mechanism that combines two complementary distributional matching terms. First, the event-rate penalty matches the marginal proportion of observed events. Second, the time-quantile penalty matches selected quantiles of the observed follow-up-time distributions. Preserving both event/censoring proportions and observed time distributions has also been emphasized in prior survival-data generation works (SurvivalGAN, Ashhad). We use these two terms alongside the global KM-based reward to act as an optimization stabilizer for the policy gradients. Because this is an empirical design choice rather than a separate methodological contribution, we do not include a dedicated ablation study.

Q2. Figure 2 "Alive": In the context of the figure, "alive" at time 24 simply denotes that the patient was right-censored at T=24T=24 without experiencing the terminal event. We will clarify this caption.

Q3. Gumbel-Softmax: For categorical variables, we employ the Gumbel-Softmax relaxation with temperature τ=0.2\tau=0.2. Thus, the generated categorical representations are differentiable soft samples rather than hard one-hot samples. These soft representations are also passed as parent inputs to downstream generators.

Q4. Gaussian Surrogate Likelihood: For continuous variables, we construct a stochastic policy using a Gaussian (Normal) distribution, where the generator output parameterizes the mean and a trainable scalar parameter determines the standard deviation.

Q5. Time Grid T\mathcal{T}: We construct the evaluation grid T\mathcal{T} using all distinct observed follow-up times in the real dataset, including both event and censoring times, sorted in ascending order: T=sort(unique({ti}i=1N))\mathcal{T}=\mathrm{sort}\left(\mathrm{unique}(\lbrace t_{i} \rbrace_{i=1}^{N})\right), where tit_{i} denotes the observed follow-up time of the (ii)-th real sample.

We hope that the responses and additional evidence address your concerns, and we kindly invite you to reconsider your overall evaluation.

Dear Reviewer kj5H,

As the Author/Reviewer discussion period concludes today, Nov. 18, we want to follow up to see if our rebuttal has successfully resolved your concerns.

We would be happy to answer any remaining questions.

If our responses have addressed your points, we would be grateful if you considered reflecting this in your overall evaluation.

Thank you again for your time and constructive feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.