{"id":"da7658bf-ece6-4f26-9b66-d9baec270bd8","arxiv_id":"2506.10140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"The authors present iSurvM, iSurvQ, and iSurvJ, nonparametric survival models that treat censored event times as imprecise interval labels and learn attention weights with kernel regression.","lead":"This paper turns survival analysis into a classification problem where censored patients are described by ranges of possible event intervals, then trains attention-weighted kernel models to predict event probabilities. It introduces three model variants that often beat the classical Beran estimator, but the advantage is not consistent across datasets and no error bars support the comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that iSurvJ/iSurvJ(G) consistently outperform Beran in accuracy is not supported by the paper's own Tables 1–2, which show several datasets where Beran wins, with no error bars and no comparably tuned baseline.","rationale":"The central claim is comparative and empirical: for it to hold, the proposed models must beat Beran reliably in accuracy and complexity. The paper's own Tables 1–2 undercut the accuracy part, showing losses on several datasets and no uncertainty quantification. The baseline also appears undertuned, so the apparent wins may reflect tuning asymmetry rather than method quality. The reader's Section 5.5 concern about the extended-interval surrogate loss is real: the justification covers only all-unique uncensored cases, yet the modification is applied universally and its bias is unanalyzed. However, that concern is secondary to the empirical comparison because even a biased training loss could in principle yield accurate test-time survival predictions. The one concrete check I propose—tuned Beran plus paired bootstrap confidence intervals—would settle whether 'consistently' is defensible. The paper does provide public code and synthetic experiments, which is genuine evidence, but those do not replace a rigorous head-to-head on the real benchmarks. The verdict should therefore remain conditional: the framework is plausible, but the central comparative claim is not yet established as stated.","tokens_in":19782,"tokens_out":7634,"duration_ms":90919,"concrete_test":"Rerun all 11 real-data comparisons with a properly tuned Beran estimator, using the same nested cross-validation and Optuna-style bandwidth selection given to the proposed models, and report paired bootstrap 95% confidence intervals for the C-index and IBS differences on every dataset. If iSurvJ/iSurvJ(G) do not win on a clear majority of datasets with statistically non-negligible margins, the claim must be softened from 'consistently outperform' to 'can outperform on some datasets after tuning'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's load-bearing claim is empirical: iSurvJ and its Gaussian variant iSurvJ(G) 'consistently outperform the Beran estimator from the accuracy and computational complexity points of view.' The paper's own tables do not support the accuracy half. In Table 1, iSurvJ has lower C-index than Beran on AIDS (0.7139 vs 0.7529) and Breast Cancer (0.6487 vs 0.6519), while iSurvJ(G) loses on 5 of 11 datasets (Veterans, AIDS, GBSG2, LND, GCD). Table 2 shows a similar pattern for the Brier score: iSurvJ(G) is worse than Beran on 6 of 11 datasets. No confidence intervals, standard errors, or significance tests are reported, so even apparent wins cannot be distinguished from noise, especially on small datasets such as LND (n=87) and GCD (n=90). The comparison also appears asymmetric: proposed models receive Optuna-based hyperparameter tuning, while no comparable tuning procedure for the Beran estimator is described. The computational-complexity half of the claim is asserted without any runtime experiments or asymptotic analysis. Because 'consistently outperform' is the headline contribution, a comparison that loses on several datasets and lacks uncertainty quantification cannot carry it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces three survival models, iSurvM, iSurvQ, and iSurvJ (plus a Gaussian-kernel variant iSurvJ(G)), built on a discrete-time, imprecise-probability view of censored survival data. Censored observations are represented as interval-valued probability distributions over time intervals, and predictions are obtained through Nadaraya-Watson-style attention weights trained by one of three loss strategies: mean over Monte Carlo-generated distributions, quantile selection of worst-case losses, or joint learning of probabilities and attention. The central claim is that these models, especially iSurvJ and iSurvJ(G), 'consistently outperform' the Beran estimator in both accuracy and computational complexity. The paper reports C-index and Brier-score comparisons on 11 real datasets and several synthetic benchmarks, and includes public code.","tokens_in":20180,"tokens_out":3685,"duration_ms":45700,"significance":"If the claims were established, the paper would contribute a novel synthesis of imprecise probability and attention-based kernel smoothing for survival analysis, with no parametric assumptions and with an explicit interval-valued representation of censoring uncertainty. The availability of public code and the breadth of datasets are positive features. However, the headline empirical claim is not supported by the paper's own tables: iSurvJ loses to Beran on several datasets, iSurvJ(G) loses on a substantial fraction of comparisons, and no uncertainty estimates or significance tests are provided. The computational-complexity claim is asserted rather than measured. Because the central contribution is framed as consistent superiority over Beran, the current evidence is insufficient for the stated conclusions.","major_comments":[{"comment":"The abstract and Section 6 claim that the proposed models 'consistently outperform' the Beran estimator, but Tables 1 and 2 show multiple losses. For example, in Table 1 iSurvJ has a lower C-index than Beran on AIDS (0.7139 vs. 0.7529) and on Breast Cancer (0.6487 vs. 0.6519), while iSurvJ(G) is worse than Beran on Veterans, AIDS, GBSG2, LND, and GCD. In Table 2, iSurvJ(G) has a worse Brier score than Beran on 6 of 11 datasets, and iSurvJ is worse on AIDS, Breast Cancer, and LND. No standard deviations, confidence intervals, or significance tests are reported, so even the apparent wins cannot be distinguished from noise, particularly on small datasets such as LND (n=87) and GCD (n=90). The accuracy half of the headline claim is therefore not supported by the presented evidence.","section":"Section 6, Tables 1-2"},{"comment":"The replacement of the exact event-interval loss by a sum over 2k+1 neighboring intervals is justified by the argument that, when all event times are unique and uncensored, the self-attention mask removes the only training example that can contribute to p_c(i), making the exact loss zero. That argument does not apply to datasets containing censored observations or tied event intervals, where other training examples contribute mass to the correct-interval probability. The modified loss is nevertheless applied to all models and all datasets, and no analysis of the resulting bias is provided. Since this surrogate loss is used in every accuracy comparison, the experimental results rest on a learning signal whose validity is not established for the data actually used.","section":"Section 5.5, Eq. (29)"},{"comment":"The comparison between the proposed models and the Beran estimator is asymmetric. The proposed models receive Optuna-based Bayesian hyperparameter optimization over embedding dimensions, learning rates, regularization coefficients, dropout, mask rates, batch sizes, and epoch counts, whereas no comparable tuning procedure for the Beran estimator is described; the only Beran parameter mentioned is a Gaussian kernel temperature in Appendix A.3.1. Observed differences could therefore reflect differences in tuning effort rather than a substantive advantage of the proposed models. The authors should either tune the Beran estimator with the same budget or justify why the chosen baseline configuration is the appropriate reference.","section":"Section 6, Paragraph on hyperparameter tuning"},{"comment":"The claim of superiority 'from the computational complexity point of view' is asserted without any runtime measurements or asymptotic analysis; no wall-clock times, flop counts, or scaling experiments are reported. In addition, the interval bounds in Appendix A.3.4 are not strong evidence for the model: because the censored interval probabilities are allowed to range over [0,1] by construction, the resulting survival-function bounds are very wide and any reasonable survival curve will lie inside them. The observation that Beran's curve falls inside the bounds should be presented as an illustration of the imprecise representation, not as an accuracy or validity finding.","section":"Section 6 and Appendix A.3.4"}],"minor_comments":[{"comment":"In the contributions list, the second model is written as 'iSurvM' twice: the text says 'survival models iSurvM, iSurvM, iSurvM' where the second and third should likely be iSurvQ and iSurvJ.","section":"Section 1, Contributions list"},{"comment":"The caption of Algorithm 2 says 'An implementation of training the model iSurvM' but the algorithm is for iSurvJ; this should be corrected.","section":"Algorithm 2 caption"},{"comment":"The C-index formula uses predicted expected event times \\widehat{T}_i and \\widehat{T}_j, but how these expected times are computed from the discrete predicted distribution is not defined anywhere in the paper.","section":"Equation (2)"},{"comment":"The Gaussian attention weight formula has a notational issue: the parameter vector is denoted w = (w_1, ..., w_n), but the denominator uses the same w_i for every term, which cannot be correct unless all weights are equal; the dependence on the target feature index should be made explicit.","section":"Equation (7)"},{"comment":"The text says the experiments 'primarily illustrate how the accuracy of the proposed model decreases as the proportion of censored data increases,' yet the abstract and Section 6 claim robustness to heavy censoring; the two statements should be reconciled.","section":"Appendix A.3.3"},{"comment":"The captions contain the typo 'censiring rate' instead of 'censoring rate'.","section":"Figures 8-9 captions"}],"recommendation":"major_revision","confidential_remarks":"The proposed framework is interesting and the code availability is a plus, but the headline claim of consistent superiority over Beran is contradicted by the paper's own tables. The authors should either significantly strengthen the empirical evaluation (tuned baseline, error bars, significance tests) and temper the abstract and conclusions, or the paper will not be publishable in its current form. I would be willing to review a revised version focusing on the modeling contribution rather than the overclaimed empirical advantage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read of the iSurv paper.\n\nThe genuinely new thing is the construction: representing censored outcomes as interval-valued probability distributions over a discrete time grid, then running kernel Nadaraya-Watson regression with trainable dot-product attention weights to produce imprecise predictive distributions. Three training schemes (mean, quantile, joint) and a Gaussian-kernel variant are laid out cleanly. This is an engineering contribution, not new theory, but it is a real one: the combination of imprecise probabilities with attention-based NW regression is not in the papers they cite. Code is public, and the synthetic experiments cover several structured regimes.\n\nWhat I do not trust is the empirical claim. The abstract says iSurvJ 'consistently outperform[s] the Beran estimator.' The paper's own Table 1 contradicts that: Beran wins on AIDS and Breast Cancer for C-index, and iSurvJ(G) loses on five of eleven datasets. Table 2 for Brier score shows a similar pattern. There are no standard deviations, no confidence intervals, no significance tests. On small datasets like LND (n=87) and GCD (n=90), the differences are within noise. The comparison is also asymmetric: the proposed models get Optuna hyperparameter optimization, while the Beran estimator is given a fixed Gaussian kernel with no comparable tuning. A kernel bandwidth matters a lot for Beran. So the central claim, as stated, is not supported by the evidence in the paper. The computational-complexity half is asserted without any runtime measurements or asymptotic analysis.\n\nThere is a second soft spot: the extended-interval loss in Section 5.5. The authors justify replacing the exact interval loss with a sum over 2k+1 neighboring intervals by arguing that, with all-unique uncensored times, self-attention masking removes the only contributing example and the exact loss is zero. That argument only holds when all uncensored event times are unique. With tied event times or censored observations, other examples contribute. The modification is applied to all models and all datasets, and the bias is never analyzed. This could distort the learning signal, and it is worth checking empirically.\n\nThe claim that Beran's curve lies inside the interval-valued bounds is also weaker than it looks: the bounds are constructed from per-censored-interval probabilities ranging over [0,1], which makes them broad by construction.\n\nGiven the mismatch between the abstract and the tables, this paper needs a serious referee, not a desk reject. The method is plausible and implemented, and the issues are addressable. The authors need to fix the baseline, add uncertainty quantification, include a modern survival baseline, and soften or remove 'consistently.'\n\nMy recommendation: send it to peer review, but expect heavy revision on the empirical side. A reader interested in imprecise-probability approaches or discrete-time survival will get something out of the framework even if the current comparisons don't hold up.","headline":"A plausible new estimator family for nonparametric survival analysis, but the headline empirical claim is contradicted by the paper's own tables and needs proper benchmarking before being taken seriously.","tokens_in":20645,"tokens_out":2024,"would_cite":false,"duration_ms":21957,"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":"The paper claims that survival analysis can be reframed as an imprecise classification problem over time intervals, encoding censored events as interval-valued probabilities and using trainable attention kernels, and that the resulting…","keywords":["survival analysis","imprecise probabilities","censored observations","attention mechanism","Nadaraya-Watson regression","discrete-time survival analysis","nonparametric estimation","trainable kernels"],"falsifier":"Train iSurvJ on a real dataset with tied uncensored event times, such as the BLCD or Rossi data, using the exact-interval loss (Eq. (28), $k=0$) instead of the windowed surrogate (Eq. (29)); if the C-index and integrated Brier score stay competitive with $k>0$, the paper's rationale for the surrogate collapses and the reported comparisons are partly driven by the modified target rather than by the imprecise representation. A more direct check: on a dataset with few ties, compare models trained with Eq. (28) and Eq. (29) and see whether the accuracy gap the paper attributes to the method disappears.","tokens_in":19583,"feed_emoji":"⏱","tokens_out":9153,"duration_ms":95887,"temperature":0.7,"pith_summary":"This paper tries to establish that time-to-event prediction can be treated as a multiclass classification problem over discrete time intervals, with censored observations represented not by a single distribution but by an interval-valued set of possible distributions. The authors propose three models — iSurvM, iSurvQ, and iSurvJ — that use Nadaraya-Watson regression with trainable attention weights as kernels, so no parametric survival model is assumed. They claim that iSurvJ and its Gaussian-kernel variant iSurvJ(G) consistently outperform the Beran estimator, a classical kernel-based nonparametric estimator of conditional survival, on both accuracy metrics and computational cost, with the advantage growing as feature dimension and censoring fraction increase. If correct, this gives a nonparametric survival method whose smoothing is learned from data rather than fixed, and whose predictions carry interval-valued uncertainty.","feed_headline":"Trainable-kernel survival models outperform the Beran estimator","feed_subtitle":"Censoring becomes interval probabilities; learned attention beats the classical kernel estimator nonparametrically.","key_machinery":"The central object is the interval-valued probability vector $\\pi^{(i)}$ over $T$ time intervals: a one-hot vector for uncensored examples and a vector with entries free in $[0,1]$ after the censoring interval for censored examples. The mechanism that carries the argument is a trainable Nadaraya-Watson weighting $p_k(x_0)=\\sum_i a_{0,i}(w)\\pi^{(i)}_k$, with attention masks preventing self-attention and dot-product or Gaussian kernels providing the weights. Training is driven by likelihood-style losses, and for iSurvJ the loss combines negative log-likelihood with entropy regularization on the learned interval probabilities. To avoid zero gradients from the self-attention mask when all uncensored event times are unique, the paper replaces the exact interval probability with a sum over $2k+1$ neighboring intervals, treating $k$ as a hyperparameter.","core_discovery":"The paper's central claim is that right-censored survival data can be encoded as imprecise multi-label classification: an uncensored event contributes a one-hot probability vector over time intervals, while a censored observation contributes the set of all probability vectors whose mass on intervals after the censoring time is free in $[0,1]$. Predictions are formed by the attention-weighted sum $p_k(x_0)=\\sum_{i=1}^N a_{0,i}(w)\\pi^{(i)}_k$, where the weights come from dot-product attention or a trainable Gaussian kernel. Three training strategies are proposed: averaging likelihood losses over random draws from the censored distributions (iSurvM), averaging only the worst quantile of those draws (iSurvQ), and jointly optimizing the interval probabilities and attention weights with an entropy regularizer (iSurvJ). The paper reports that iSurvJ and iSurvJ(G) beat the Beran estimator on most real and synthetic benchmarks, that the Beran survival curve lies inside the interval-valued survival bounds the model produces, and that the model's unconditional survival function tracks the classical nonparametric estimator.","pith_inferences":["A natural extension the authors leave implicit is to use the interval-valued survival bounds as a formal uncertainty region for decision rules, such as delaying an intervention while the upper survival bound stays high.","The same imprecise-classification encoding could be reused for interval-censored and left-censored events by changing which entries of $\\pi^{(i)}$ are free, which the paper mentions only briefly as a motivation for imprecision.","The $2k+1$ window surrogate can be viewed as label smoothing over neighboring time intervals; if that is the active ingredient, the method's gains might partly come from regularization rather than from the imprecise representation, a hypothesis the paper does not test.","Extending the models to competing risks or time-varying covariates, which the authors list as future work, would require replacing the scalar interval target with a vector-valued one but would keep the attention machinery unchanged."],"forward_implications":["The framework removes the need for parametric assumptions: the same training procedure handles linear, nonlinear, and interacting feature structures by learning attention weights from data.","Heavy censoring no longer degrades predictions as sharply as it does for the Beran estimator, so the model is a candidate for datasets where most events are unobserved.","The interval-valued survival function produced for each instance gives lower and upper bounds that can enclose the Beran estimate, offering uncertainty information alongside point predictions.","Because the number of trainable attention parameters can be chosen freely, users can trade model capacity against computational cost without changing the formulation.","Among the model's hyperparameters, the window width $k$ for uncensored intervals shows a plateau in accuracy beyond a threshold, so a moderate $k$ avoids unnecessary computation."],"supporting_citations":[{"why":"Supplies the Beran estimator, the nonparametric baseline that the proposed models must outperform in every comparison.","marker":"[20]"},{"why":"Introduces Nadaraya-Watson regression, the kernel-weighted form that the paper adapts into trainable attention weights.","marker":"[33]"},{"why":"The companion smoothing-regression formulation that justifies expressing predictions as weighted averages of interval probabilities.","marker":"[34]"},{"why":"Defines the dot-product attention that the paper trains as its key-query weighting mechanism.","marker":"[35]"},{"why":"Introduces the self-attention and masking machinery that the paper borrows for computing and restricting attention weights.","marker":"[36]"},{"why":"Provides the Dirichlet distribution used to generate random probability vectors from the interval-valued sets for censored observations in iSurvM and iSurvQ.","marker":"[61]"}],"fun_headline_variants":["Imprecise survival models beat the Beran estimator","Censoring as interval probabilities: trainable-kernel survival","iSurvJ: imprecise classification for survival data","Survival without parametric assumptions via imprecise probabilities","Nonparametric survival with trainable attention kernels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that training uncensored examples on the sum of probabilities over a window of $2k+1$ neighboring intervals, instead of on the exact probability of the true event interval, is a faithful learning target; the paper justifies this replacement only for datasets with unique uncensored event times and applies it to all datasets without analyzing the bias it may introduce for tied event times or censored data.","fun_headline_variants_meta":{"raw":{"variants":["Imprecise survival models beat the Beran estimator","Censoring as interval probabilities: trainable-kernel survival","iSurvJ: imprecise classification for survival data","Survival without parametric assumptions via imprecise probabilities","Nonparametric survival with trainable attention kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1455,"prompt_tokens":1011,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":627,"tokens_out":444,"duration_ms":5152,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:34:45.117225+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train iSurvJ on a real dataset with tied uncensored event times, such as the BLCD or Rossi data, using the exact-interval loss (Eq. (28), $k=0$) instead of the windowed surrogate (Eq. (29)); if the C-index and integrated Brier score stay competitive with $k>0$, the paper's rationale for the surrogate collapses and the reported comparisons are partly driven by the modified target rather than by the imprecise representation. A more direct check: on a dataset with few ties, compare models trained with Eq. (28) and Eq. (29) and see whether the accuracy gap the paper attributes to the method disappears.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Beran estimator, the nonparametric baseline that the proposed models must outperform in every comparison."},{"cited_title":"Nadaraya","cited_arxiv_id":null,"evidence_quote":"Introduces Nadaraya-Watson regression, the kernel-weighted form that the paper adapts into trainable attention weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The companion smoothing-regression formulation that justifies expressing predictions as weighted averages of interval probabilities."},{"cited_title":"Luong, H","cited_arxiv_id":null,"evidence_quote":"Defines the dot-product attention that the paper trains as its key-query weighting mechanism."},{"cited_title":"Vaswani, N","cited_arxiv_id":null,"evidence_quote":"Introduces the self-attention and masking machinery that the paper borrows for computing and restricting attention weights."},{"cited_title":"Rubinstein and D.P","cited_arxiv_id":null,"evidence_quote":"Provides the Dirichlet distribution used to generate random probability vectors from the interval-valued sets for censored observations in iSurvM and iSurvQ."}],"review_version":1}