{"id":"cf4d4c00-f7c4-4fc6-aa8a-48cf7835777d","arxiv_id":"2412.07638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SurvBETA, an ensemble of Beran survival estimators aggregated by three attention mechanisms, reports the highest C-index on 8 of 12 survival benchmarks when its simplified linear-programming training variant is used.","lead":"SurvBETA is a new survival-modeling method that builds an ensemble of Beran kernel estimators and combines their survival predictions with attention weights. It aims at censored medical or tabular data where complex cluster structures make standard Cox-based models less reliable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Significance claim in Table 2 collapses under multiple-comparison correction; SurvBETA-with-opt no longer clearly beats RSF/GBM Cox/GBM AFT.","rationale":"The strongest claim is not about the model's internal mechanics but about its empirical superiority. The construction of the special-case optimization is a methodological contribution, and the experiments are extensive (12 datasets, 100 repetitions, public code). However, the statistical analysis that converts these experiments into a superiority claim is flawed. The p-values in Table 2 are computed from paired t-tests across 12 datasets, but the paper does not account for multiple comparisons. With five baselines, the family-wise error rate is inflated; after a simple Bonferroni correction, three of the five comparisons lose significance. The lack of error bars in Table 1 compounds the issue: the reader cannot tell whether the reported C-index differences are stable across repetitions. This is more load-bearing than the reader's identified weakest assumption (that only the special case of the attention model is implemented) because the paper explicitly frames the experiments as evaluating the special case, and the conclusion only claims the special case works. The special-case issue is a scope limitation, not a fatal flaw. If the corrected p-values fail, the conclusion must be weakened from 'significantly outperforms all' to 'competitive, with best C-index on 8 of 12 datasets,' which is a meaningful downgrade. Therefore the verdict should remain CONDITIONAL (or UNCHANGED given the reader already assigned CONDITIONAL), with the condition being the corrected statistical analysis.","tokens_in":17632,"tokens_out":7215,"duration_ms":63251,"concrete_test":"Recompute Table 2 by applying a Bonferroni correction for the five pairwise comparisons (multiply each p-value by 5, cap at 1). If any corrected p-value for the comparisons against RSF, GBM Cox, or GBM AFT exceeds 0.05, then the paper's claim that SurvBETA with optimization significantly outperforms all considered models is not supported. (A stronger alternative is to run a Friedman test with Nemenyi post-hoc across the 12 datasets, which is the standard method for multiple model comparisons.)","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline empirical claim—that SurvBETA with optimization 'significantly outperforms all the considered models'—rests entirely on the p-values in Table 2. These p-values come from a paired t-test over 12 datasets, with no correction for the fact that five pairwise comparisons are made. Applying a Bonferroni correction (×5) to the reported p-values yields 0.00035, 0.00035, 0.175, 0.19, and 0.095 for Beran, SurvBETA-without-opt, RSF, GBM Cox, and GBM AFT respectively. Thus the comparisons against RSF, GBM Cox, and GBM AFT are no longer significant at the 0.05 level, so the 'all models' conclusion is not supported. Additionally, Table 1 reports only mean C-indices over 100 random splits, without standard deviations or confidence intervals, so it is impossible to assess whether the numerical gaps (e.g., 0.7254 vs 0.7001 for Veterans) are within run-to-run noise. The mathematical derivation of the special-case optimization appears coherent, and the code is released; the issue is specifically the statistical evidence for superiority.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SurvBETA, an ensemble method in which Beran estimators serve as weak learners whose conditional survival functions are aggregated by attention weights. Three attention mechanisms are described: the Beran kernel weights, instance-dependent prototypes of bootstrap subsamples, and a global aggregation weight. A general trainable formulation is given, followed by a simplified special case based on Huber's epsilon-contamination model, where the aggregation weights take the form (1-epsilon)*softmax + epsilon*v and training reduces to a linear or quadratic programming problem. The paper reports synthetic experiments studying ensemble size, cluster distance, subsample size, and other factors, and real-data experiments on 12 benchmarks comparing SurvBETA variants with a single Beran estimator, RSF, GBM Cox, and GBM AFT in terms of C-index, with pairwise significance tests.","tokens_in":17915,"tokens_out":8453,"duration_ms":82315,"significance":"If the empirical claims survive proper statistical analysis, the paper offers a useful and relatively simple ensemble construction for survival analysis that is trainable through a convex reformulation, and it is one of the few attempts to build ensembles around Beran estimators. Strengths of the work include a publicly released implementation, an explicit derivation of the LP/QP special case, and synthetic experiments specifically designed to probe cluster structure, where the proposed method behaves sensibly. The main weakness is that the headline empirical claim, that SurvBETA with optimization 'significantly outperforms all the considered models', is not supported by the statistical evidence as presented.","major_comments":[{"comment":"The claim that SurvBETA with optimization 'significantly outperforms all the considered models' rests on five pairwise p-values from a paired t-test over 12 datasets, with no multiple-comparison correction. Applying a Bonferroni correction for the five comparisons in the first row of Table 2 turns 0.035, 0.038, and 0.019 for RSF, GBM Cox, and GBM AFT into 0.175, 0.19, and 0.095, none of which is significant at the 0.05 level. The paper should either declare the comparison family and control its error rate, use a test designed for comparisons across multiple datasets such as the Wilcoxon signed-rank test recommended by Demsar, or substantially soften the conclusion. As written, the 'all models' claim is not supported by the reported evidence.","section":"Section 6.2, Table 2"},{"comment":"Table 1 reports only mean C-indices over 100 random 60/20/20 splits, with no standard deviations or confidence intervals. Consequently, it is impossible to assess whether individual advantages such as 0.7254 versus 0.7001 on Veterans, or 0.7499 versus 0.6387 on WPBC, are stable across splits or within run-to-run noise. Please report standard errors or confidence intervals for every entry, and if the significance analysis in Table 2 is based on the 12 dataset-level means, account for the variance across the 100 splits as well, for example through a hierarchical or permutation-based comparison.","section":"Table 1"},{"comment":"The authors state in Section 6 that only the simplified epsilon-contamination special case of Eq. (28) is implemented and that the general attention model of Eqs. (24)-(27) is not considered because of the large number of training parameters, and Section 7 repeats that the general problem was not implemented. The abstract and title nevertheless present SurvBETA as a model 'using three attention mechanisms'. The empirical validation therefore covers only a special case. Please add a prominent scope statement that the benchmarked system is the simplified special case, and ensure that the abstract, conclusion, and title do not imply that the full three-attention architecture has been empirically validated.","section":"Section 6 and Section 7"}],"minor_comments":[{"comment":"In the definition of the aggregation attention weight, the second factor in the softmax uses x_j, although the weight is supposed to depend on the prototype e(A_k,x); please correct the notation so the argument matches Eq. (13).","section":"Eq. (26)"},{"comment":"The kernels are defined with u = -||x-x_i||^2/tau, but the Epanechnikov, triangular, and quartic expressions use u^2 or |u|, so the minus sign is confusing; please define u as a nonnegative distance or clarify the intended convention.","section":"Section 6.1, kernel definitions"},{"comment":"In the definition of L_MAE, the index m and the notation S_l(x_i) are not defined; this appears to be a different aggregation from Eq. (18) and should be clarified.","section":"Eq. (23)"},{"comment":"After the substitution beta_k = epsilon*v_k, the optimization problem lists v among the variables, but v does not appear in the objective or constraints; either remove v from the variable list or explicitly state that beta_k encodes the trainable attention weights.","section":"Eq. (47)"},{"comment":"References [26] and [27] appear to be the same paper by Shi Hu et al.; please merge or distinguish them.","section":"References"},{"comment":"The caption contains a typo: 'comprison' should be 'comparison'.","section":"Table 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The statistical analysis in Tables 1-2 is the main obstacle to acceptance. If the authors add a proper multiple-comparison correction or reframe the empirical claim, and clarify the scope of the evaluated model, the paper would be within reach of acceptance. The special case is closely related to the authors' earlier work [56,57]; the discussion should more explicitly state what is new relative to those papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the construction is genuinely new: Beran estimators as weak learners, attention at three levels (within the estimator, for prototypes, and for aggregating), and a nearest-neighbor bootstrap that keeps subsamples intersecting. The LP/QP reformulation of the epsilon-contamination special case is internally consistent, and the code and 12 real benchmark experiments are public. Second, the headline empirical claim—that SurvBETA with optimization \"significantly outperforms all the considered models\"—does not survive a multiple-comparison correction. The Table 2 p-values are raw paired t outcomes, and after Bonferroni (×5) the comparisons to RSF, GBM Cox, and GBM AFT lose significance at 0.05. Table 1 reports only mean C-indices over the 100 splits, with no standard deviations, so even the raw gaps (e.g., Veterans 0.7254 vs 0.7001) cannot be separated from run-to-run noise.\n\nWhat is genuinely good: the method is a coherent new ensemble paradigm, not a patch on an existing forest. The K-nearest-neighbor bootstrap with overlapping subsamples is a thoughtful answer to the problem that fully disjoint subsamples make prototypes unhelpful. The synthetic experiments (distance between clusters, number of estimators) isolate a real strength: the ensemble holds up when the single Beran estimator degrades. The paper is also transparent that the general scaled-dot-product attention in Eqs. (24)-(26) is not implemented; Section 6 states only the simplified special case is considered. That is a credit, not a defect, though it means the title's \"SurvBETA\" should be read as that special case.\n\nThe soft spots are real but focused. The significance test is the load-bearing flaw. The authors follow Demsar's paired t-test but do not correct for the five comparisons, and then over-interpret the raw p-values. Fixing this is straightforward: either apply a correction and soften the conclusion to \"outperforms the Beran estimator and the no-opt variant, and is competitive with RSF/GBM,\" or report confidence intervals in Table 1. A second, milder issue: the LM AE term in the QP includes absolute values but the variable ψi handles that correctly; no problem there. The citation pattern is fine; [56,57] are self-citations but they are directly prior work, and the central claim doesn't depend on them being strong.\n\nWho this is for: researchers working on survival ensembles and attention-based tabular models. It deserves a serious referee, not a desk reject. I would send it out but with the expectation that the statistical evidence be tightened before acceptance. The method may well be right; the paper just hasn't shown it is better than RSF and GBM at the 0.05 level once you look at all comparisons.","headline":"Genuinely new ensemble construction with public code, but the headline significance claim collapses under multiple-comparison correction.","tokens_in":18410,"tokens_out":3431,"would_cite":false,"duration_ms":29169,"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":"SurvBETA: an attention-weighted ensemble of Beran estimators for survival analysis.","keywords":["survival analysis","Beran estimator","ensemble learning","attention mechanism","epsilon-contamination model","censored data","C-index","bootstrap subsampling"],"falsifier":"Run the public SurvBETA code on the same twelve datasets with ordinary random bootstrap subsamples instead of the intersecting K-nearest-neighbor scheme; if the reported C-index advantage over random survival forests and gradient-boosted Cox disappears, then the load-bearing component is the sample-generation design, not the attention-weighted aggregation.","tokens_in":17458,"feed_emoji":"📊","tokens_out":6041,"duration_ms":57523,"temperature":0.7,"pith_summary":"SurvBETA is a proposed ensemble for survival analysis in which each weak learner is a Beran estimator, a kernel-based rule that estimates the chance of surviving past a given time while using the proximity of other patients' feature vectors to the analyzed patient. Individual survival-function predictions are combined not by simple averaging but by attention weights that measure how close the analyzed patient is to a prototype of each bootstrap subsample. Training the aggregation weights in full generality is hard, so the authors derive a simplified special case from the epsilon-contamination model, which turns training into a linear program over the ensemble weights. On synthetic data with two separated clusters, the model keeps its accuracy as the gap between clusters grows, where a single Beran estimator degrades. On twelve real survival datasets, the optimized variant achieves the best C-index on eight of them, and the reported paired t-test p-values are below 0.05 against all compared baselines.","feed_headline":"Attention-weighted survival ensemble wins 8 of 12 benchmarks","feed_subtitle":"Kernel-based weak learners plus three attention levels win most real-data comparisons in this study.","key_machinery":"The central object is the Beran estimator, written as a product over event times of one minus a normalized kernel weight, which estimates conditional survival functions nonparametrically. Around it the paper builds a three-level attention mechanism: alpha weights within each Beran estimator, mu weights that form a Nadaraya-Watson prototype of each bootstrap subsample relative to the test instance, and gamma weights that aggregate the M subsample survival functions. The special case that makes training tractable replaces gamma with (1-epsilon)*softmax + epsilon*v, where v is a probability vector over subsamples; substituting this into a C-index surrogate with hinge loss produces a linear program in v and slack variables. The bootstrap procedure itself is also load-bearing: subsamples are K-nearest-neighbor balls around randomly chosen centers so that they intersect, ensuring several aggregation weights stay comparable for any test point.","core_discovery":"The paper's central claim is that the Beran estimator, usually studied as a standalone nonparametric estimator, can serve as a weak learner in an attention-based ensemble, and that the resulting model is competitive with or better than established survival ensembles. The mechanism rests on three attention levels: kernel weights inside each Beran estimator, prototype weights that locate each bootstrap subsample relative to a test instance, and global aggregation weights that combine the survival functions. Because the aggregation weights depend on how far the test point is from a subsample's prototype, the ensemble can discount weak learners trained mostly on distant clusters. The full dot-product attention version is stated but deliberately not implemented; what is implemented and tested is the special case where each aggregation weight is a mixture of a softmax distance term and a training parameter, derived from the epsilon-contamination model, which makes optimization linear. With that special case, the paper reports the best C-index on most of twelve real datasets and p-values below 0.05 against the single Beran estimator, random survival forests, gradient-boosted Cox, and gradient-boosted AFT models.","pith_inferences":["The authors' cluster explanation suggests a testable prediction the paper does not test: the advantage of SurvBETA over Cox-type models should grow with cluster separation and shrink on homogeneous data; a reader could verify this on non-clustered benchmarks.","The full dot-product attention version remains unimplemented; if the simplified weighting is truly a good surrogate, then a regularized version of the full form might improve further, but that is speculation beyond the paper.","Because the simplified aggregation is linear in v, the method could be extended to online or streaming survival settings where ensemble weights are updated as new censored observations arrive; the paper does not discuss this."],"forward_implications":["If the claims hold, attention-weighted Beran ensembles are a viable alternative for tabular survival data, particularly when covariates form separated clusters.","The architecture separates weak learners from the two outer attention levels, so the Beran estimators could be swapped for other survival models without changing the aggregation scheme.","The epsilon-contamination simplification means training ensemble weights reduces to a linear program, avoiding the overfitting that the authors attribute to the full dot-product attention form.","The reported results imply that models built on Cox proportional hazards can be outperformed on clustered data by kernel-based local ensembles.","Even without the trained contamination weights, the attention-based aggregation improves on a single Beran estimator, so part of the gain comes from the ensemble geometry alone."],"supporting_citations":[{"why":"Supplies the Beran estimator, the kernel-based weak learner that every ensemble member uses.","marker":"[5]"},{"why":"Supplies the epsilon-contamination model that motivates the simplified attention-weight form.","marker":"[28]"},{"why":"Supplies the simplified attention-weight definition and the prototype-based aggregation idea this model adapts to survival.","marker":"[56]"},{"why":"Defines random survival forests, one of the main baselines the model must beat.","marker":"[29]"},{"why":"Defines the gradient-boosted Cox baseline used in the comparisons.","marker":"[46]"},{"why":"Defines the gradient-boosted accelerated-failure-time baseline used in the comparisons.","marker":"[2]"},{"why":"Defines the C-index, the concordance measure used both as a loss and as the reported performance metric.","marker":"[22]"},{"why":"Provides the paired t-test procedure used to claim statistical significance over the baselines.","marker":"[11]"}],"fun_headline_variants":["SurvBETA: three attention levels, one winning ensemble","Beran-based survival ensemble with attention wins most","Attention-weighted kernel ensemble edges out RSF and Cox","Survival ensemble with three attention levels wins most benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluated model is not the full attention architecture but the simplified epsilon-contamination special case, and the paper's empirical claims assume that this simplified weighting, together with the intersecting K-nearest-neighbor bootstrap, preserves the behaviour of the full three-attention model on real data.","fun_headline_variants_meta":{"raw":{"variants":["SurvBETA: three attention levels, one winning ensemble","Beran-based survival ensemble with attention wins most","Attention-weighted kernel ensemble edges out RSF and Cox","Survival ensemble with three attention levels wins most benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000512,"raw_usage":{"total_tokens":2514,"prompt_tokens":992,"completion_tokens":1522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1458}},"tokens_in":608,"tokens_out":1522,"duration_ms":11618,"temperature":1.0,"reasoning_tokens":1458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:51.524033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the public SurvBETA code on the same twelve datasets with ordinary random bootstrap subsamples instead of the intersecting K-nearest-neighbor scheme; if the reported C-index advantage over random survival forests and gradient-boosted Cox disappears, then the load-bearing component is the sample-generation design, not the attention-weighted aggregation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the paired t-test procedure used to claim statistical significance over the baselines."}],"review_version":1}