{"id":"6d4f93a4-20d2-4eb4-bcfe-bc9752907628","arxiv_id":"2506.01348","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Wasserstein distributionally robust Cox model, solved via a conic program, improves survival prediction accuracy under covariate shift and outliers.","lead":"The paper proposes a version of Cox survival regression that is trained to be robust when the data distribution shifts or contains outliers. It replaces the usual Cox loss with a distributionally robust version and tests it on a heart-attack dataset and a miscarriage dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The implemented DRL-Cox uses a gamma=3 truncation of Eq. (5)'s O(N^2) constraints; dropping constraints relaxes the feasible set, so the tested model is not the model for which Theorem 2 and Lemma 5 guarantees are proved.","rationale":"I read the paper as claiming a theoretically grounded robust Cox variant: Theorem 2 gives a tractable upper bound for (4), exact when X = R^d, and Lemma 5 gives a finite-sample guarantee. The most load-bearing point is not the surrogate loss, which the authors disclose in Section 2.3 and which can reasonably be viewed as adding the event itself to its risk set; it is the relation between the guaranteed program (5) and the program executed in Section 4. The gamma = 3 truncation is introduced as a computational necessity with only a one-sentence assertion of integrity. Mathematically, dropping constraints in a minimization relaxes the feasible set, so the optimal value can only decrease; hence the upper-bound and worst-case guarantee may fail for the model whose metrics are reported. This is verifiable by running the full program. I also note a printing error in Eq. (5): the inner sum uses sum_{k=1}^i with k also the constraint index, while the proof's s_i definition requires the risk-set sum to run over the constraint index k; the code should be checked against the proof. These concerns do not refute the overall approach, but they mean the central claim is conditional on showing that the truncated program either preserves the upper bound or is replaced by the full program in the experiments. This aligns with the reader's CONDITIONAL verdict, though through a different mechanism than the loss-surrogate concern; hence partial agreement.","tokens_in":13883,"tokens_out":37110,"duration_ms":403169,"concrete_test":"Re-run the WHAS500 distributional-shift experiments (Figure 1) with the full constraint set of Eq. (5), i.e., gamma = N, on the same train/validation folds and the same epsilon-selection rule used for the reported gamma = 3 runs. If the full-program objective or the resulting C-index/iAUC differs materially from the gamma = 3 results, or if the gamma = 3 objective is below the full-program objective beyond solver tolerance, the truncation is not benign and the empirical claim is not established for the model actually tested.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing issue is the gap between the model for which guarantees are proved and the model actually solved in experiments. Eq. (5) is an O(N^2) exponential conic program in which, for each i, s_i must lower-bound the maximum over k >= i of the expression log(exp(beta'x_i) + risk-set sum) - beta'x_i - alpha(y_i - y_k). Section 4 then replaces this by keeping only the constraints with k < i + gamma, gamma = 3. Dropping inequality constraints in a minimization relaxes the feasible set, so the gamma = 3 program's optimal value can be strictly below the optimal value of (5). Theorem 2's upper bound and Lemma 5's finite-sample guarantee apply to (5), not to this relaxed variant; the model evaluated in Tables 1-3 and Figure 1 is therefore not the model whose robustness is claimed. The paper's assertion that the truncation 'maintains the integrity of the optimization process' is unsupported: the active maximum over k can occur at a risk set far beyond i + 2, especially under distributional shift where later risk sets have different composition. The reported gains could thus be an artifact of the extra relaxation rather than of the Wasserstein DRL construction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRL-Cox, a distributionally robust version of Cox regression. It replaces the standard Cox partial-likelihood loss with a modified individual loss (Eq. 3), embeds this loss in a Wasserstein-ambiguity min-max problem (Eq. 4), and derives an exponential conic program (Eq. 5) claimed to be an upper bound on the robust loss, with equality claimed when the covariate space is R^d. The paper also states a finite-sample performance guarantee (Lemma 5). Experiments on the WHAS500 dataset and the PRESTO miscarriage cohort compare DRL-Cox with regular and penalized Cox models, AFT, and RSF under covariate shifts and outliers, reporting higher C-index and iAUC for DRL-Cox. Code and partial data are made available.","tokens_in":14189,"tokens_out":8745,"duration_ms":90279,"significance":"If the derivation and finite-sample guarantee were fully rigorous, the paper would offer a useful tractable formulation for robust survival analysis and a principled alternative to sample-splitting DRL approaches. The empirical setup is relevant, and the authors provide code and use both a public benchmark and a large cohort. However, the current proof has unverified minimax steps, the implemented model differs from the analyzed program by truncating constraints, and the loss is a surrogate for the Cox partial likelihood; these gaps currently limit the strength of the central claims.","major_comments":[{"comment":"The step from Eq. (16a) to Eq. (16b) exchanges sup over theta and inf over (r, alpha) before taking sup over x, and the later step immediately before Eq. (19) exchanges max over j and inf over (theta, alpha). Both steps are minimax interchanges, but the paper invokes the Minimax Theorem without checking the compactness, convexity, or saddle-point conditions for these specific functions, which are not shown to be jointly convex-concave in the relevant variables. Because these interchanges are the core of the derivation of Eq. (5), Theorem 2 is not fully established as written.","section":"Sec. 3.1, Eq. (16b)"},{"comment":"The claim that Eq. (5) is an exact reformulation of Eq. (4) when X = R^d is asserted but not proved. The argument shows that one inequality (Eq. (16c)) is tight when X = R^d, but inequalities remain in Eqs. (18) and (19), and no argument demonstrates that those become equalities on R^d. The exactness claim should either be proved with the missing steps or removed and replaced by a statement of the conditions under which equality holds.","section":"Sec. 3.1, Theorem 2"},{"comment":"The experiments solve a gamma = 3 truncated version of Eq. (5) that keeps only constraints with k < i + gamma. Removing inequality constraints from a minimization relaxes the feasible set, so the truncated program's optimal value can be strictly smaller than the optimal value of Eq. (5). The upper-bound statement in Theorem 2 and the finite-sample guarantee in Lemma 5 apply to Eq. (5), not to this truncated model, and the unsupported assertion that the truncation 'maintains the integrity of the optimization process' does not address the gap. The reported improvements could therefore be artifacts of the extra relaxation rather than of the Wasserstein DRL construction.","section":"Sec. 4, constraint truncation"},{"comment":"The individual loss in Eq. (3) is not the Cox partial-likelihood loss; it inserts an additional e^{beta'x} term inside the log. All theoretical results and experiments use this modified loss, so the finite-sample guarantee and robustness claims pertain to a surrogate objective rather than to the standard Cox model. The statement that this modification 'preserves theoretical consistency with the proportional hazards assumption' is not substantiated, and the manuscript should either justify the surrogate as a principled loss or soften the claims connecting it to standard Cox regression.","section":"Sec. 2.3, Eq. (3)"},{"comment":"Lemma 5 is quoted from Zhao and Guan without proof and assumes a bounded and 'finite' data space, which is ambiguous. The guarantee requires epsilon >= B(alpha), but in the experiments epsilon is selected by cross-validation and a concentration method, with no demonstration that the selected radii satisfy the lemma's threshold. Thus the reported empirical results are not covered by the stated finite-sample guarantee.","section":"Sec. 3.2, Lemma 5 and Sec. 4, choice of epsilon"}],"minor_comments":[{"comment":"The constraint in Eq. (5) uses the index i both for the constraint and for the summation index inside the log; a distinct summation index, such as sum_{m=1}^k, would remove the ambiguity.","section":"Sec. 3.1, Eq. (5)"},{"comment":"Tables 2 and 3 report averages over five iterations without confidence intervals or significance tests, making it difficult to assess whether the reported advantages of DRL-Cox are meaningful.","section":"Sec. 4, Tables 2 and 3"},{"comment":"Figure 1 would benefit from error bars or confidence bands, especially since the text emphasizes non-monotonic behavior across shift levels.","section":"Sec. 4.1, Figure 1"},{"comment":"The phrase 'the data space X x Y is bounded and finite' is ambiguous: boundedness and finiteness are different conditions, and the lemma should specify which is intended.","section":"Sec. 3.2, Lemma 5"},{"comment":"The text refers to the 'training loss c_hat_N' but only J_hat_N is defined; the notation should be introduced consistently before it is used.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the venue and addresses a relevant problem. The main concern for the editor is that the theoretical claims are ahead of what is proved: the minimax interchanges are not justified, the X = R^d exactness is asserted rather than demonstrated, and the experiments solve a truncated program that is not the one covered by the guarantees. These issues are addressable in principle, which is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll be direct: the paper's main idea is worth a look—Wasserstein DRL applied to a Cox-type loss with an exponential conic reformulation and a finite-sample bound is a natural combination that, to my knowledge, hasn't been done in survival analysis. The authors are honest about modifying the individual loss (Eq. 3) and they provide code and data for reproducibility. The empirical comparison includes reasonable baselines.\n\nThe soft spots are real. Most serious is the constraint truncation in Section 4. Equation (5) has O(N^2) constraints; the experiments use only k < i + 3 (gamma=3). Dropping inequality constraints in a minimization relaxes the feasible set, so the tested program can have a lower optimal value than (5). Theorem 2's upper bound and Lemma 5's finite-sample guarantee apply to (5), not to the relaxed variant. The one-sentence assertion that truncation 'maintains the integrity' is unsupported; under distributional shift the active maximum could easily lie beyond i+2. Until the authors either solve the full program for moderate N or prove that the truncated constraint set preserves the bound, the theoretical claims don't cover the reported results.\n\nSecond, the surrogate loss deviates from the Cox partial likelihood. The authors disclose this, which helps, but it means the theory and the name 'Cox' apply to a modified objective. The practical impact may be small, but it should be stated more prominently.\n\nThird, the proof of Theorem 2 has gaps: the minimax interchange in (16b) is invoked without checking the required compactness or convexity, the X=R^d equality is asserted rather than derived, and Lemma 5 is imported without verification. The result is likely true, but the paper doesn't fully demonstrate it.\n\nFinally, the empirical claims are overstated. The reported gains over regular Cox are about 1–2% absolute C-index, not the 'more than 5%' in the introduction unless one compares against the weakest baseline. No error bars or standard deviations are given, and the hyperparameter selection for epsilon is described loosely.\n\nOverall, the core idea has merit and the authors are transparent about their choices, but the gap between theory and implementation is load-bearing. I would send it to peer review because the combination is new and the authors can likely address the gamma issue, but I would not use it as-is as a basis for my own work. If the full program is intractable, they need a rigorous argument that the truncation is harmless, or the guarantees must be re-proven for the truncated set.","headline":"Plausible DRL-Cox extension, but the implemented constraint truncation breaks the theoretical guarantees, and the surrogate loss is not the Cox partial likelihood.","tokens_in":14678,"tokens_out":4471,"would_cite":false,"duration_ms":47075,"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":"A distributionally robust Cox model reduces to an exponential conic program and outperforms standard survival models under distributional shift and outliers.","keywords":["distributionally robust optimization","Cox proportional hazards model","survival analysis","Wasserstein distance","ambiguity set","exponential conic programming","finite-sample guarantee","censored data"],"falsifier":"On a small synthetic dataset with a bounded feature space, evaluate both sides of Theorem 2 by solving the min-max problem (4) with a fine grid search and solving the conic program (5); a strict gap between the two values would show the upper bound is loose in exactly the bounded-feature regime where the real-data experiments are run.","tokens_in":13684,"feed_emoji":"🛡️","tokens_out":16205,"duration_ms":142800,"temperature":0.7,"pith_summary":"This paper sets out to establish that Cox regression can be made robust to distributional shift and outliers by optimizing the worst-case loss over a Wasserstein ball of distributions around the training data. To make that min-max problem tractable, it replaces the per-subject Cox loss with a surrogate that adds an extra exponential term, so the loss becomes convex in the covariates and nonnegative for long-duration subjects. The main theorem converts the resulting robust problem into an exponential conic program that is an upper bound on the original min-max value and is exact when the feature space is all of $\\mathbb{R}^d$. A finite-sample guarantee says that with an ambiguity radius $\\epsilon$ chosen like $\\sqrt{\\log(1/\\alpha)/N}$, the true expected loss stays below the robust training objective with probability at least $1-\\alpha$. If these claims are correct, survival models could retain their accuracy when the data used for training and the data seen in practice come from different distributions.","feed_headline":"Worst-case Cox model beats standard survival models under shift","feed_subtitle":"Wasserstein-ball training keeps survival rankings accurate even when covariates shift or outliers contaminate the data.","key_machinery":"The central object is the Wasserstein-distance ambiguity set $\\Omega_\\epsilon = \\{P : W_s(P, \\hat{P}_N) \\le \\epsilon\\}$ centered at the empirical distribution, coupled with the surrogate individual loss (3) that inserts an extra $e^{\\beta'x}$ term into the log and treats all training subjects as fixed constants. That insertion makes the loss convex in the covariate direction and nonnegative for long-duration subjects, which is what allows Wasserstein duality, the dual norm, and the convex conjugate to convert the min-max problem into the exponential conic program (5). The proof also discretizes the duration coordinate, whose piecewise-constant, non-convex behavior blocks a direct duality argument.","core_discovery":"On the paper's own terms, the discovery is that the distributionally robust Cox problem (4), built on the redefined individual loss (3), has a tractable convex reformulation: the exponential conic program (5), whose objective is $\\epsilon\\|(\\beta,\\alpha)\\|_q$ plus an averaged weighted log-sum-exp loss over subjects, with $O(N^2)$ constraints relating the slack variables $s_i$ to risk-set sums. The proof passes through Wasserstein strong duality, discretizes the non-convex duration direction, and evaluates the convex conjugate of the surrogate loss, which pins down the penalty coefficient as $\\|\\beta\\|_q$. The relaxation is tight when the covariate space is exactly $\\mathbb{R}^d$. The paper further claims a finite-sample robustness guarantee, and reports that in simulations and on two healthcare datasets the model consistently beats standard Cox, penalized Cox, accelerated failure time, and random survival forest baselines in C-index and time-dependent AUC under contamination.","pith_inferences":["The paper's guarantee and the conic reformulation are stated for the surrogate loss (3), not for the standard Cox partial likelihood (1); if a reader cares about the classical Cox objective, the connection is asserted by analogy rather than proven.","Because the conic program is exact only when the feature space is all of $\\mathbb{R}^d$, on bounded feature spaces the program is an upper bound and the reported gains may partly reflect that looseness.","A clean-data comparison of the surrogate loss against standard Cox would isolate the robustness contribution from the regularizing effect of the extra exponential term and the $\\epsilon\\|(\\beta,\\alpha)\\|_q$ penalty.","The same duality-and-discretization template could be applied to other losses that are convex in features but non-convex in one coordinate, such as ranking or quantile objectives."],"forward_implications":["The robust Cox objective is solvable with exponential conic programming, and its $O(N^2)$ constraint set can be truncated to $O(\\gamma N)$ constraints; the paper uses $\\gamma=3$ and reports about thirty seconds per trial, so the method is practical on moderate datasets.","Choosing the ambiguity radius $\\epsilon$ at or above $B(\\alpha)$, which scales as $\\sqrt{\\log(1/\\alpha)/N}$, yields a finite-sample guarantee that the true expected loss is bounded by the empirical robust objective with probability at least $1-\\alpha$.","If the surrogate loss is accepted as the objective, DRL-Cox should keep its ranking accuracy when covariates shift or outliers contaminate features, and the experiments consistently show C-index and iAUC gains over regular and penalized Cox, sample-splitting Cox, AFT, and random survival forests.","Because the derivation needs convexity in the covariate direction, the same route does not directly extend to deep survival models; the authors point to constraining weight matrices as a possible adaptation."],"supporting_citations":[{"why":"It defines the Wasserstein-distance ambiguity set and the measure-concentration method used to select the ambiguity radius.","marker":"Chen and Paschalidis (2020)"},{"why":"It supplies the data-driven Wasserstein distributionally robust optimization template and its tractable-reformulation and performance-guarantee results, which the paper builds on.","marker":"Esfahani and Kuhn (2018)"},{"why":"It gives the strong dual of the inner supremum over the Wasserstein ball, the first step of the proof of Theorem 2.","marker":"Gao and Kleywegt, 2023, Cor.2"},{"why":"It is the source of the finite-sample performance guarantee in Lemma 5 that bounds the true expected loss by the empirical DRL objective.","marker":"Zhao and Guan, 2015, Prop. 3"},{"why":"It defines the proportional-hazards partial-likelihood loss that the paper modifies into the surrogate loss (3).","marker":"Cox (1972)"},{"why":"It is the prior distributionally robust Cox model using sample splitting that motivates the surrogate-loss construction and serves as a comparison baseline.","marker":"Hu and Chen (2022)"},{"why":"It provides the preconception cohort from which the real-world miscarriage dataset used in the outlier experiments is drawn.","marker":"Wise et al. (2015)"},{"why":"It provides the WHAS500 dataset used in the covariate distributional-shift experiments.","marker":"Goldberg et al. (2000)"}],"fun_headline_variants":["Wasserstein-robust Cox survives shift and beats baselines","DRL-Cox: robust survival under distribution shift","Ambiguity sets make Cox survival models shift-proof","Fast robust Cox via exponential conic programming","Survive data shift: Wasserstein-robust Cox wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the redefined individual loss in Eq. (3), which adds an extra exponential term and treats all training subjects as fixed constants, is an acceptable surrogate for the Cox partial likelihood; the paper itself acknowledges in Section 2.3 that this design deviates from the standard Cox loss, and if the surrogate is not faithful, the theoretical bound and the name 'Cox' do not transfer to the standard model.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein-robust Cox survives shift and beats baselines","DRL-Cox: robust survival under distribution shift","Ambiguity sets make Cox survival models shift-proof","Fast robust Cox via exponential conic programming","Survive data shift: Wasserstein-robust Cox wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3449,"prompt_tokens":865,"completion_tokens":2584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":2504}},"tokens_in":481,"tokens_out":2584,"duration_ms":18800,"temperature":1.0,"reasoning_tokens":2504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:44:48.992352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small synthetic dataset with a bounded feature space, evaluate both sides of Theorem 2 by solving the min-max problem (4) with a fine grid search and solving the conic program (5); a strict gap between the two values would show the upper bound is loose in exactly the bounded-feature regime where the real-data experiments are run.","supporting_citations":[],"review_version":1}