{"id":"45240a5b-3c0d-4d88-adb0-8285b40b0d66","arxiv_id":"2509.22352","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SurvDiff is an end-to-end diffusion model that jointly generates covariates, event times, and right-censoring indicators for survival data, guided by a Cox-based survival loss.","lead":"This paper introduces SurvDiff, a diffusion model that generates synthetic patient data for survival analysis, including event times and censoring flags. It is designed to help researchers share or augment clinical datasets without exposing real patient records.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core claim that the Cox survival loss 'preserves the censoring mechanism' is unsupported: Eq. 10 supervises only risk scores from denoised covariates and never models the censoring distribution; Table 7 shows mixed survival fidelity.","rationale":"The reader's verdict is CONDITIONAL, and my concern supports that rather than moving the verdict to reject: the method is plausible, TabDiff-style diffusion with T and E as generated dimensions could well capture much of the joint distribution, and the empirical tables show genuine gains on some metrics. The problem is specifically that the paper's headline guarantee about censoring is not derived from the objective and is not measured. Eq. 10 is a ranking loss over denoised covariates; it does not contain a term for P(E=0 | X, T), for the censoring time, or for the generated T/E. Thus the claim that SurvDiff 'ensures' preservation of the censoring mechanism overstates what the loss can do. The appendix's own Table 7 provides additional evidence: on several rows a baseline beats SurvDiff, and the appendix text says 'broadly comparable performance,' so the abstract's 'consistently outperforms' is not supported by the complete reported results. I am not accusing the authors of anything; the failure mode is a standard mismatch between an auxiliary training loss and a downstream distributional guarantee. The proposed test is directly executable: compare reverse KM curves and conditional censoring rates. If those checks pass, the central claim would be substantially strengthened; if they fail, the contribution would need to be reframed as a heuristic regularizer rather than a censoring-preserving objective. Keeping the paper CONDITIONAL with this additional required evaluation is the right call.","tokens_in":16991,"tokens_out":7506,"duration_ms":66384,"concrete_test":"Release code and the missing hyperparameters (tau, alpha, lambda_max), then for each of the three datasets compare real and synthetic censoring distributions directly: (1) compute reverse Kaplan-Meier estimates of the censoring-time distribution from SurvDiff, TabDiff, and SurvivalGAN outputs and report the Kolmogorov-Smirnov distance to the real data; (2) fit a logistic regression of E=0 on covariates (or stratified empirical censoring rates by age quartiles and treatment) and report standardized differences between real and synthetic coefficients or rates. If SurvDiff is not closer to the real censoring distribution than both baselines, or if ablating L_surv leaves these censoring-specific metrics unchanged, then the claim that the survival loss preserves the censoring mechanism is unsupported and should be withdrawn from the abstract.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SurvDiff's central contribution, stated in the abstract and Section 1, is that the survival-tailored loss (i) reproduces event-time distributions and (ii) preserves the censoring mechanism. The only survival-specific training signal is L_surv in Eq. 10: a weighted Cox partial likelihood computed on risk scores r_i = f_theta(x_i), where x_i are denoised covariates at random diffusion timesteps and the risk sets use observed event times t_i. A Cox partial likelihood is constructed to estimate hazard ratios without modeling the censoring distribution; under standard independent-censoring assumptions it is valid for any censoring mechanism. It therefore cannot, by itself, 'encode' or 'preserve' that mechanism. Moreover, L_surv does not take the generated event time or event indicator as input, so its gradients do not directly shape the generated (T, E); those are supervised only by the generic diffusion terms Lcont and Ldisc, which a plain tabular diffusion baseline also has. The paper gives no argument or measurement connecting L_surv to the censoring distribution. The survival metrics in Supplement D and Table 7 are event-time summaries: KM MSE and RMST gap. They do not compare censoring times or conditional censoring probabilities, and the paper itself notes RMST gap can mask compensating deviations. Table 7 is internally mixed: on AIDS, TabDiff has lower RMST gap (0.0066 vs 0.0134) and SurvivalGAN has lower KM MSE (0.0002 vs 0.0003); on GBSG2, TabDiff has lower KM MSE (0.0005 vs 0.0062). Supplement D describes the results as 'broadly comparable performance,' which contradicts the abstract's 'consistently outperforms.' Because the method's claimed novelty is survival-tailored censoring preservation, this missing evaluation and theoretical mismatch are load-bearing. Missing hyperparameters (tau, alpha, lambda_max) additionally prevent replication of the exact objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SurvDiff, a diffusion model for generating synthetic survival data. It jointly generates mixed-type covariates, event times, and a right-censoring indicator via a masked-diffusion process for discrete variables and a Gaussian diffusion process for continuous variables, with an added Cox partial-likelihood loss on risk scores computed from denoised covariates. The authors claim that this survival-tailored loss reproduces event-time distributions and preserves the censoring mechanism, and they report experiments on AIDS, GBSG2, and METABRIC comparing against NFlow, TVAE, CTGAN, TabDiff, and SurvivalGAN. The main empirical claims are that SurvDiff consistently outperforms baselines in covariate fidelity and downstream survival metrics.","tokens_in":17492,"tokens_out":3653,"duration_ms":28731,"significance":"If the empirical claims held, SurvDiff would be a useful contribution as the first diffusion model tailored to synthetic survival data, with the practical benefit of end-to-end joint generation. The paper includes a broad benchmark across three medical datasets, five generative baselines, five downstream survival models, and several fidelity metrics, and it reports means and standard deviations over ten seeds. The idea of augmenting a diffusion objective with a Cox-based loss is plausible and relevant. However, the significance is currently limited by the gap between the strong abstract claim of consistent outperformance and the mixed results in the tables, and by the unsupported claim that the Cox loss preserves the censoring mechanism.","major_comments":[{"comment":"The abstract claims that SurvDiff 'consistently outperforms state-of-the-art generative baselines' in distributional fidelity and survival metrics; this is contradicted by the reported tables. In Table 2, TVAE has a lower Wasserstein distance on AIDS (0.0779 vs 0.0869) and TabDiff has a lower Wasserstein distance on METABRIC (0.0539 vs 0.0547); in Table 3, TabDiff has a slightly higher C-index on METABRIC (0.6118 vs 0.6117); and in Table 7, TabDiff has a lower RMST gap on AIDS (0.0066 vs 0.0134), SurvivalGAN has a lower KM MSE on AIDS (0.0002 vs 0.0003), and TabDiff has a lower KM MSE on GBSG2 (0.0005 vs 0.0062). The claim of consistent outperformance should be qualified, or the experiments and abstract should be revised to match the observed results.","section":"Abstract and Section 5.1, Table 3"},{"comment":"The central claim that the survival-tailored loss 'preserves the censoring mechanism' is not supported by the formulation. L_surv in Eq. (10) is a Cox partial likelihood on risk scores r_i computed from denoised covariates, with risk sets based on observed times; it does not model the censoring distribution, and a Cox partial likelihood is valid under independent censoring regardless of the censoring mechanism, so it cannot by itself encode or preserve that mechanism. The generated event time T and event indicator E are supervised only by the generic diffusion terms Lcont and Ldisc, which a plain tabular diffusion baseline also has. To support the claim, the paper should either provide a mechanism by which L_surv shapes the censoring distribution or report direct tests of censoring (e.g., censoring-time distributions or conditional censoring probabilities).","section":"Section 4.3, Eq. (10)"},{"comment":"The survival-fidelity evidence is mixed and the paper itself acknowledges that the RMST gap can mask compensating deviations. In Table 7, SurvDiff does not dominate: on AIDS, TabDiff has a lower RMST gap (0.0066 vs 0.0134) and SurvivalGAN has a lower KM MSE (0.0002 vs 0.0003); on GBSG2, TabDiff has a lower KM MSE (0.0005 vs 0.0062). The paper describes these results as 'broadly comparable performance,' which is inconsistent with the abstract's 'consistently outperforms.' The authors should either temper the contribution statement or add additional survival metrics that directly assess the censoring distribution.","section":"Supplement D, Table 7"}],"minor_comments":[{"comment":"The text refers to 'Figure 4.3' when describing the three components; this should be 'Figure 2'.","section":"Section 4, Figure 2"},{"comment":"The sentence 'SurvDiff generating reliable synthetic datasets' is missing a verb; it should be 'SurvDiff generates reliable synthetic datasets.'","section":"Section 6, Conclusion"},{"comment":"Several hyperparameter entries lack a space, e.g., 'No. Epochs4000', 'No. Iterations1500'; this should be fixed for readability.","section":"Appendix B, Tables 4 and 6"},{"comment":"The abstract contains a duplicated phrase: 'in both distributional fidelity and survival model evaluation metrics across multiple medical datasets' and later 'across multiple datasets'; consider removing one occurrence.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an arXiv preprint and does not state code availability; the authors are encouraged to release code if possible. The paper's scope fits cs.LG and ML for health. I recommend that the editor seek the additional analyses requested in the major comments, especially a direct assessment of the censoring distribution, before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good to talk. Quick take on SurvDiff: it is a genuinely useful application paper, not a methodological breakthrough, and its central rhetorical claim—that the Cox auxiliary loss preserves the censoring mechanism—is not actually supported by the equations or the experiments. The authors should soften the abstract or add a direct test.\n\nWhat's new: it is the first end-to-end diffusion model for synthetic survival data, jointly generating mixed covariates, event time, and right-censoring indicator. The architecture builds on TabDiff-style masked/Gaussian diffusion, and adds a weighted Cox partial likelihood as auxiliary loss plus a sparsity-aware weighting scheme. The empirical work is real: three medical datasets, multiple baselines including SurvivalGAN, ten seeds, and TSTR downstream evaluation with five survival models. On covariate JS distance and most downstream metrics, SurvDiff comes out ahead, especially on AIDS/GBSG2 where censoring is strong. The small-sample sensitivity analysis is a nice addition.\n\nNow the soft spots. The abstract says 'consistently outperforms' but the tables don't support that without qualification. TVAE beats it on Wasserstein for AIDS, TabDiff on METABRIC for Wasserstein and C-index; in Supplement D TabDiff is better on several survival metrics. Supplement D itself says 'broadly comparable,' which contradicts the abstract. That's a real overclaim, but a fixable one: just report the numbers honestly and frame it as state-of-the-art on average.\n\nThe bigger issue is the censoring mechanism claim. The only survival-specific loss is L_surv, a Cox partial likelihood on risk scores from denoised covariates. Cox partial likelihood is designed to estimate hazard ratios while making no assumptions about censoring distribution; it does not model or constrain that distribution. The generated censoring indicators and event times are supervised only by the generic diffusion losses. So the assertion that SurvDiff 'preserves the censoring mechanism' is unsupported by the method's math. The paper also never directly evaluates censoring—no comparison of censoring distributions, no conditional censoring probabilities. The KM and RMST metrics in Table 7 are event-time summaries that can mask compensating differences. This is load-bearing because the paper's distinctiveness is precisely the survival-tailored censoring preservation. I'd want a direct evaluation—for instance, compare conditional censoring rates given covariates, or at least the marginal censoring distribution—before the claim stays in the abstract.\n\nReproducibility: the text uses tau, alpha, lambda_max in Eqs. 10–13 but the hyperparameter table omits them, and there's no code. That's fixable and should be mandatory.\n\nBottom line: this is a solid, useful contribution to synthetic survival data generation. It deserves peer review, but the authors need to correct the overclaims, add a direct censoring evaluation, and release code and full hyperparameters. I'd be inclined to accept after major revision.","headline":"Solid application paper that overclaims: the Cox auxiliary loss does not by itself preserve censoring, and several tables contradict the 'consistently outperforms' framing.","tokens_in":17954,"tokens_out":1907,"would_cite":true,"duration_ms":16599,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SurvDiff, an end-to-end diffusion model, jointly generates covariates, event times, and right-censoring and, the paper reports, outperforms existing generators on fidelity and downstream survival metrics across three medical datasets.","keywords":["survival analysis","synthetic data generation","diffusion models","right-censoring","time-to-event data","tabular data","Cox proportional hazards","clinical machine learning"],"falsifier":"On a simulated survival dataset with known event and censoring time generators, fit SurvDiff and compare the conditional distribution of the censoring indicator and censoring times in synthetic versus real samples; any material divergence—for example, a covariate that predicts censoring in the truth but not in the synthetic data—would falsify the censoring-preservation claim.","tokens_in":16746,"feed_emoji":"🩺","tokens_out":10135,"duration_ms":87485,"temperature":0.7,"pith_summary":"Survival data are hard to synthesize because event times are often right-censored, and a synthetic cohort must reproduce both the covariate distribution and the censoring mechanism. SurvDiff is a diffusion model that generates covariates, event times, and the censoring indicator together in one end-to-end pass, with an added loss term that ranks patients by survival risk using a weighted Cox partial likelihood. On the AIDS, GBSG2, and METABRIC medical datasets, the paper reports that SurvDiff matches real covariate distributions more closely than existing tabular and survival generators and yields better or equal downstream survival-model performance, with the largest gains on heavily censored data. The paper positions SurvDiff as the first end-to-end diffusion model designed for generating synthetic survival data.","feed_headline":"Diffusion model preserves censoring in synthetic survival data","feed_subtitle":"Synthetic cohorts retain event-time structure and let survival models train without the original patient data.","key_machinery":"The central object is the combined objective $L_{\\text{total}} = L_{\\text{diff}} + \\lambda_{\\text{surv}} L_{\\text{surv}}$. SurvDiff uses a variance-exploding Gaussian diffusion for continuous covariates and event times, plus a masked multinomial diffusion for discrete covariates and the event indicator, following the TabDiff architecture. A denoising transformer reconstructs original values from noise, and a small survival head maps the denoised covariates to a scalar risk score. The survival loss is a weighted Cox partial negative log-likelihood in which only observed events contribute to the numerator while censored patients remain in the risk sets, with an exponential decay weight for late events to reduce instability from small risk sets; an adaptively scaled $\\lambda_{\\text{surv}}$ balances this loss against the diffusion reconstruction loss.","core_discovery":"The central claim is that a single diffusion model can jointly generate mixed-type covariates, event times, and right-censoring indicators, and that a survival-tailored loss—an extension of the Cox partial likelihood with exponential decay weighting that downweights late events—is what makes censoring-aware generation work. The paper contrasts this end-to-end design with SurvivalGAN, which generates covariates first and then samples event times conditionally, arguing that the staged approach propagates errors and loses censoring fidelity. In experiments across three clinical cohorts, SurvDiff attains lower Jensen–Shannon and Wasserstein distances for covariates, better C-indices and Brier scores in train-on-synthetic/test-on-real evaluation, and competitive Kaplan–Meier and RMST metrics; the authors conclude that the survival loss and joint generation are responsible for the gains, not the diffusion backbone alone.","pith_inferences":["A direct test the paper does not report is whether the synthetic censoring-time distribution matches the real one: comparing censoring Kaplan-Meier curves or the covariate-conditional censoring rate would settle the strongest claim more cleanly than the reported downstream metrics.","Because the survival head is built on a Cox partial likelihood, SurvDiff's generated event-time structure may inherit the proportional-hazards assumption; on data with crossing hazards, a non-proportional survival head would be a natural variant to test.","The same joint-generation recipe—diffusion reconstruction loss plus a loss that scores what downstream models will predict—appears portable to competing risks or interval censoring, since only the outcome head and risk-set definition would change."],"forward_implications":["Survival models trained on SurvDiff's synthetic cohorts and tested on real patients achieve better C-index and Brier scores than models trained on other generators' synthetic data, the paper's main downstream test.","The largest gains occur on the most heavily censored datasets, so the survival-tailored loss is doing the work rather than the diffusion backbone alone.","Joint end-to-end generation avoids the error propagation of staged generators, so covariates, event times, and censoring indicators come from one model instead of separately fitted pieces.","In downsampled cohorts, SurvDiff's advantage over TabDiff grows, indicating the method transfers to small-sample clinical settings."],"supporting_citations":[{"why":"Defines the survival-data generation problem and the staged SurvivalGAN baseline that SurvDiff is designed to surpass.","marker":"Norcliffe et al., 2023"},{"why":"Supplies the mixed-type tabular diffusion architecture and the power-mean and log-linear noise schedules that SurvDiff adapts, and serves as the TabDiff baseline.","marker":"Shi et al., 2024b"},{"why":"Provides the partial likelihood that the survival-tailored loss extends with censoring-aware weights.","marker":"Cox, 1972"},{"why":"Gives the Cox negative log-likelihood formulation and survival-head design used in DeepSurv-style risk scoring.","marker":"Katzman et al., 2018"},{"why":"Supplies the score-based stochastic differential equation formulation and variance-exploding forward process used for continuous covariates and event times.","marker":"Song et al., 2021"},{"why":"Supplies masked diffusion over discrete states, used for the discrete covariates and the event indicator.","marker":"Sahoo et al., 2024"},{"why":"Defines the concordance index, the paper's primary downstream ranking metric.","marker":"Harrell et al., 1982"},{"why":"Defines the Brier score, the calibration metric used in downstream survival evaluation.","marker":"Brier, 1950"}],"fun_headline_variants":["SurvDiff: joint diffusion for events, censoring, and covariates","Survival data synthesis that respects the censoring mechanism","One diffusion pass to generate realistic synthetic survival data","Diffusion model matches event-time and censoring in fake cohorts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the assumption that a survival-risk loss computed on noisy, partially denoised versions of the data teaches the generator to reproduce the real censoring mechanism, even though no part of the training objective directly checks that mechanism.","fun_headline_variants_meta":{"raw":{"variants":["SurvDiff: joint diffusion for events, censoring, and covariates","Survival data synthesis that respects the censoring mechanism","One diffusion pass to generate realistic synthetic survival data","Diffusion model matches event-time and censoring in fake cohorts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1396,"prompt_tokens":918,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":534,"tokens_out":478,"duration_ms":4496,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:43:42.916057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a simulated survival dataset with known event and censoring time generators, fit SurvDiff and compare the conditional distribution of the censoring indicator and censoring times in synthetic versus real samples; any material divergence—for example, a covariate that predicts censoring in the truth but not in the synthetic data—would falsify the censoring-preservation claim.","supporting_citations":[{"cited_title":"SurvivalGAN : Generating time-to-event data for survival analysis","cited_arxiv_id":null,"evidence_quote":"Defines the survival-data generation problem and the staged SurvivalGAN baseline that SurvDiff is designed to surpass."},{"cited_title":"Score-based generative modeling through stochastic differential equations","cited_arxiv_id":null,"evidence_quote":"Supplies the score-based stochastic differential equation formulation and variance-exploding forward process used for continuous covariates and event times."},{"cited_title":"Chiu, Alexander M","cited_arxiv_id":null,"evidence_quote":"Supplies masked diffusion over discrete states, used for the discrete covariates and the event indicator."}],"review_version":2}