{"id":"fd9be6a9-1831-433b-85b8-337da62390c0","arxiv_id":"2506.08313","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A new parametric lifetime distribution called EEPHND is proposed and reported to reach 0.9997 concordance on lung cancer data, but the evaluation is circular.","lead":"This paper introduces EEPHND, a new six-parameter mixture of an exponentiated exponential Pareto distribution and a half-normal distribution for survival data. It claims near-perfect concordance on a lung cancer dataset, but that result rests on an in-sample comparison that is not a valid predictive test.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported CI 0.9997 is not evidence of predictive accuracy: EEPHND has no covariates, so the fitted survival function is common to all patients and any in-sample concordance computed from S(t_i) is near 1 by construction.","rationale":"The most load-bearing assumption in the empirical section is that the reported C-index measures predictive accuracy. It cannot, because the model has no covariates; all fitted survival curves are identical. The near-perfect CI is a mathematical consequence of evaluating a decreasing function at ordered event times, not a sign of model quality. This is fatal to the strongest claim in the abstract. I focus on this rather than the censoring-adjusted likelihood (which the reader also flagged) because even if the likelihood were correctly censoring-adjusted, the CI would remain meaningless; the predictive claim would still fail. The distributional construction, moments, and sampling algorithm are standard mixture mechanics and are not the source of the problem. The rejection is warranted, and I see no reason to change the reader's verdict.","tokens_in":10773,"tokens_out":4625,"duration_ms":51644,"concrete_test":"Split the lung dataset into training and test sets (e.g., 70/30). Fit EEPHND on the training set using the paper's MLE procedure (Eq. 32) and obtain S_hat(t). On the test set, assign every patient the same predicted risk score, e.g., S_hat(t*) at a fixed horizon t* equal to the median follow-up time, and compute Harrell's concordance index; it will be 0.5, demonstrating no patient-level discrimination. Also recompute CI using the paper's implicit score S_hat(t_i) and show analytically that it is forced to approximately 1 for any decreasing survival function. If the reported 0.9997 cannot be reproduced as an out-of-sample discrimination measure, the predictive superiority claim is void.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that EEPHND achieves a Concordance Index of 0.9997 on the lung cancer data, outperforming Cox PH (0.6029) and Kaplan-Meier (0.9982). This claim is not interpretable as predictive accuracy because the fitted EEPHND is a single parametric survival distribution with no covariates (Section 3.2.1, Table 2). Every patient therefore receives the same predicted survival curve S_hat(t) = 1 - F_hat(t). A concordance index requires a risk score that varies across patients. The paper never defines the score used to compute CI. If the score is S_hat(t_i) evaluated at each patient's own observed time, then because S_hat is strictly decreasing in t, the score is a deterministic decreasing function of the observed time; Harrell's C for any strictly decreasing survival function evaluated this way is approximately 1, regardless of whether the model fits the data. This explains why the no-covariate Kaplan-Meier estimator also achieves CI 0.9982. The near-perfect CI is therefore an artifact of the evaluation protocol, not evidence of real predictive discrimination. Out-of-sample, the model cannot distinguish patients: at any fixed horizon, it assigns identical survival probabilities to all subjects. The predictive-accuracy component of the headline claim collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new parametric lifetime distribution, the Exponentiated-Exponential-Pareto-HalfNormal Mixture (EEPHND), defined as a two-component mixture of an exponentiated exponential-Pareto distribution and a half-normal distribution. The authors derive closed-form expressions for the density, CDF, survival, hazard, odds, moments, and moment generating function, and present a maximum likelihood estimation procedure. They validate the model on simulated data and on the lifelines lung cancer dataset, reporting that EEPHND attains a concordance index (CI) of 0.9997, outperforming Cox proportional hazards and Kaplan-Meier. The manuscript's central empirical claim is that EEPHND offers superior predictive accuracy for survival data while retaining a closed-form parametric structure.","tokens_in":11054,"tokens_out":3628,"duration_ms":44528,"significance":"If the claims were correct, a flexible closed-form parametric mixture with a near-perfect concordance index on a standard lung cancer dataset would be a notable contribution, because it could serve as a tractable alternative to Cox or Kaplan-Meier baselines for population-level survival modeling. The paper does provide some useful algebraic groundwork: the mixture construction, the sampling algorithm, and the simulation-based demonstration of flexibility are all coherent in outline, and the theoretical development of the moments is clearly intended. However, the central statistical contributions are not currently supported. The likelihood in Eq. (32) ignores right censoring, the density in Eq. (2a) is inconsistent with the CDF in Eq. (4), and the reported CI of 0.9997 is an in-sample artifact of a model without covariates, not a measure of predictive discrimination. These issues invalidate the headline empirical claims, so the manuscript cannot be accepted in its present form.","major_comments":[{"comment":"The PDF of the EEP component in Eq. (2a) is inconsistent with the CDF in Eq. (4). Differentiating F(x) = p1[1 - (1 - e^{-λ(x/β)^θ})^α] + p2 erf(x/(σ√2)) with respect to x gives a term proportional to αλθ/β (x/β)^{θ-1} e^{-λ(x/β)^θ} (1 - e^{-λ(x/β)^θ})^{α-1}. Equation (2a), however, contains the factor [1 - (1 - e^{-λ(x/β)^θ})^{α-1}], which is not the derivative of the stated CDF. This error propagates into the likelihood (Eq. 31), the hazard, the reliability, and the odds function in Sections 2.6–2.8, so the entire distributional framework is not mathematically self-consistent.","section":"Equations (2a), (4), and Sections 2.6–2.7"},{"comment":"The log-likelihood in Eq. (32) is the sum of log densities for all observations, with no adjustment for censoring. The real lung cancer dataset (Section 3.2.1) contains a censoring indicator, and a valid likelihood must be Σ[δ_i ln f(t_i) + (1-δ_i) ln S(t_i)]. As written, the MLE is only justified for complete data, and the fitted parameters reported in Table 2 (α=0.0001, θ=0.01, etc.) are not valid estimates for the right-censored lung cancer data. Because the subsequent AIC/BIC/CAIC comparisons in Table 3 and the survival estimates at t=0.012 depend on these parameters, the real-data results are unsupported.","section":"Section 2.10, Eq. (32)"},{"comment":"The reported concordance index of 0.9997 is not a meaningful measure of predictive accuracy. EEPHND has no covariates, so the fitted model assigns the same survival function S_hat(t) to every patient; there is no patient-specific risk score. If the CI is computed by ranking patients by S_hat(t_i) at their own observed times, then, because S_hat is strictly decreasing in t, the ranking is a deterministic function of the observed survival time. Any such monotone survival function will produce a concordance index near 1, which explains why Kaplan-Meier also attains CI=0.9982. Out-of-sample, the model cannot distinguish between patients at any fixed horizon. The paper must define the score used for CI and, at minimum, evaluate predictive performance on a hold-out set or via cross-validation.","section":"Section 3.2.2, Table 2"},{"comment":"The survival function is algebraically incorrect. The paper writes R(x) = 1 - [p1(1 - (1 - (1 - e^{-λ(x/β)^θ})^α)) + p2 erf(x/(σ√2))]. Since the CDF in Eq. (4) is p1(1 - (1 - e^{-λ(x/β)^θ})^α) + p2 erf(...), the correct survival function is 1 - p1(1 - (1 - e^{-λ(x/β)^θ})^α) - p2 erf(...). The extra complement inside the p1 term changes the functional form and invalidates the derived hazard, odds, and any results that use R(x).","section":"Section 2.6"},{"comment":"The skewness and kurtosis formulas contain a factual substitution error: Eq. (8) correctly gives E(X^r) = p1(β/(αλ)^{1/θ})^r Γ(r/θ + 1) + p2(...). However, Eq. (12) and the subsequent skewness and kurtosis expressions replace β/(αλ)^{1/θ} with β/√(θαλ), which changes the scale factor and is not algebraically equivalent. Consequently, the reported skewness and kurtosis coefficients are not the moments of the defined EEPHND distribution. This also propagates to the moment generating function in Section 2.5.","section":"Section 2.3 and 2.4"}],"minor_comments":[{"comment":"The text states that survival times were 'rescaled to the range [0,1]' but does not justify this transformation or explain how it affects the likelihood, the parameters, or the interpretation of the survival estimates. Since the data are right-censored, the rescaling should be described in detail.","section":"Section 3.2.1"},{"comment":"Equation numbering jumps from Eq. (20) to Eq. (31) without any equations 21–30. This appears to be a formatting error, but it makes it difficult to follow the development of the likelihood.","section":"Section 2.10"},{"comment":"The inverse-transform sampling formula for the EEP component in Eq. (19) is X = β[-1/λ ln(1 - (1 - V)^{1/α})]^{1/θ}, which matches the CDF only if the EEP CDF is as in Eq. (4). Given the inconsistency between Eq. (2a) and Eq. (4), the sampling algorithm needs to be checked against the corrected density.","section":"Section 2.9"},{"comment":"Several references are numbered inconsistently (e.g., [12] is cited both for the EEPD and for a different work in the reference list), and the reference list contains items not clearly cited in the text. The authors should carefully proofread the bibliography and the equation numbering before resubmission.","section":"General"}],"recommendation":"reject","confidential_remarks":"The manuscript has multiple load-bearing statistical errors: the likelihood ignores censoring, the PDF is inconsistent with the CDF, and the reported CI is an in-sample tautology for a model without covariates. These are not local presentation issues; they invalidate the real-data analysis. The authors would need to substantially revise the statistical framework and redo the empirical evaluation for the paper to become viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a routine new parametric mixture, and the headline result is an artifact. The model has no covariates, so the fitted survival curve is the same for every patient; computing a concordance index from S(t_i) gives near 1 by construction. The Kaplan-Meier CI of 0.9982 in the same table is the tell. The comparison to Cox (0.6029) is apples-to-oranges, since Cox uses patient covariates and EEPHND does not.\n\nWhat is genuinely new is the specific mixture of the exponentiated-exponential-Pareto (really exponentiated Weibull/Pareto-type) component and the half-normal. The authors derive moments, hazard, reliability, odds function, a sampling scheme, and an MLE section. The mixture construction and the algebra are standard textbook work, but they are done in one place. That is a legitimate, if modest, contribution.\n\nNow the soft spots, in order of severity. First, the concordance index is computed in-sample from a monotone survival function common to all patients. No risk score is defined. If the score is S_hat(t_i), then for any i<j with t_i < t_j, S_hat(t_i) > S_hat(t_j), so every comparable pair is concordant. The 0.9997 is tautological, not predictive. Second, the likelihood in Eq. (32) is a product of densities with no censoring term. The lung cancer data are right-censored; ignoring that makes the MLE unjustified and the parameter estimates suspect. Third, the PDF in Eq. (2a) is inconsistent with the CDF in Eq. (4) and with the moments in Eq. (8): the factor should be (1 - e^{-λ(x/β)^θ})^{α-1}, not 1 minus that. As printed, the density for α=1 is identically zero. That is not a typo in a minor section; it is the centerpiece of the model, and anyone implementing from the paper would produce a wrong density.\n\nMinor issues: no code or data are provided, and the \"heavy tail\" flexibility is asserted but never demonstrated against real data. The bootstrap CIs in the simulation section are fine but do not transfer to the real-data analysis.\n\nWho is this for? A reader building a catalog of mixture distributions might want to note the construction, but the empirical claims should be ignored. The paper does not deserve a serious referee in its current form. If the authors fix the PDF, add proper censoring to the likelihood, and drop the in-sample CI as evidence, a revised version could be a modest contribution to the distribution-flexibility literature. As submitted, desk reject.","headline":"A routine new mixture distribution whose headline CI is a no-covariate in-sample artifact; the empirical claims collapse, but the mixture construction is a legitimate small contribution.","tokens_in":11652,"tokens_out":4929,"would_cite":false,"duration_ms":51768,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62E10","62N01","62F10","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes the EEPHND, a five-parameter mixture of an Exponentiated-Exponential-Pareto distribution and a Half-Normal distribution, and claims it reaches a concordance index of 0.9997 on a lung cancer survival dataset…","keywords":["EEPHND","mixture distribution","Exponentiated-Exponential-Pareto","Half-Normal","survival analysis","concordance index","lung cancer","maximum likelihood"],"falsifier":"Refit the model to the lung cancer data with a censoring-adjusted likelihood and evaluate the concordance index on held-out data, for example via time-dependent AUC with inverse-censoring weights; if the index falls toward the Cox level (near 0.6) or the fitted parameters shift sharply, the headline 0.9997 is an artifact of in-sample fitting.","tokens_in":10557,"feed_emoji":"🫁","tokens_out":5453,"duration_ms":64812,"temperature":0.7,"pith_summary":"This paper proposes a new parametric lifetime distribution, EEPHND, built as a mixture of an Exponentiated-Exponential-Pareto component and a Half-Normal component. The authors aim to show that this model captures both early-time symmetric behavior and long-tail behavior while keeping closed-form expressions for the density, survival, hazard, moments, and reliability functions. They derive the distribution's properties, estimate parameters by maximum likelihood, validate on simulated data, and apply the model to a real lung cancer survival dataset. Their central empirical claim is that EEPHND fits that dataset better than competing parametric models by AIC, BIC, and CAIC, and achieves a concordance index of 0.9997, above the Cox proportional hazards model and the Kaplan-Meier estimator. If the claim holds, the model offers a fully parametric, closed-form alternative for survival and reliability modeling where classical distributions are too rigid.","feed_headline":"New mixture distribution beats Cox and Kaplan-Meier on lung cancer data","feed_subtitle":"EEPHND wraps a heavy-tailed exponential-Pareto component around a half-normal core and reports a concordance index of 0.9997.","key_machinery":"The EEPHND density is the finite mixture $f(x) = p_1 f_{\\mathrm{EEPD}}(x;\\alpha,\\beta,\\theta,\\lambda) + p_2 f_{\\mathrm{HN}}(x;\\sigma)$, where the EEPD component is the exponentiated exponential-Pareto family with shape and scale parameters $\\alpha, \\beta, \\theta, \\lambda$, and the Half-Normal component has scale $\\sigma$. Because the mixture is linear, the cumulative distribution function, survival function $S(x) = 1 - F(x)$, and hazard $h(x) = f(x)/S(x)$ all stay in closed form. That closed form is what carries the argument: it yields explicit raw moments, central moments, skewness, kurtosis, a moment generating function, reliability expressions, and a simple inverse-transform sampling scheme. Maximum likelihood estimation uses Newton-Raphson to solve the score equations numerically.","core_discovery":"The paper's central discovery is a tractable mixture distribution whose survival function stays closed-form. The density is $f(x) = p_1 f_{\\mathrm{EEPD}}(x;\\alpha,\\beta,\\theta,\\lambda) + p_2 f_{\\mathrm{HN}}(x;\\sigma)$, combining the heavy-tailed Exponentiated-Exponential-Pareto component with the symmetric Half-Normal component, with the mixture weight $p_1$ governing the balance between early and late risk. The authors report that on simulated data the model tracks the empirical CDF closely, and on the lung cancer dataset it produces the lowest AIC, BIC, and CAIC among EEPD, Log-Normal, Gamma-Rayleigh, and Half-Normal. They also report a concordance index of 0.9997, which they interpret as superior predictive accuracy relative to Cox PH (0.6029) and Kaplan-Meier (0.9982). This is the first time, the paper argues, that a fully parametric mixture with closed-form survival and hazard functions matches or exceeds a nonparametric survival benchmark on this kind of data.","pith_inferences":["The fitted real-data mixture weight $p_1 \\approx 0.01$ suggests the Half-Normal component dominates the lung cancer fit; testing the same dataset against a simpler Half-Normal-only model would clarify whether the heavy-tailed EEP component earns its additional parameters.","Comparing EEPHND's concordance index directly to Kaplan-Meier is conceptually unusual, since Kaplan-Meier is not a per-subject prediction rule; a fairer comparison would use censoring-adjusted time-dependent AUC or Brier scores on held-out subjects.","The closed-form survival function could be extended to incorporate covariates through a proportional-hazards or accelerated-failure-time link, which would combine the model's flexible baseline with the covariate interpretability currently provided only by Cox PH.","The reported $α = 0.0001$ estimate on real data indicates potential parameter instability; repeating the fit with profile likelihood or penalized estimation would show how much of the CI result depends on a single extreme parameter value."],"forward_implications":["EEPHND provides closed-form density, CDF, survival, hazard, moments, and reliability functions, making it directly usable for parametric survival and reliability analysis without numerical integration.","On simulated data, the model matches the empirical CDF better than a normal model, and on the lung cancer dataset it reports the lowest AIC, BIC, and CAIC among the compared parametric models.","If the reported concordance index holds, a fully parametric mixture can match or slightly beat the nonparametric Kaplan-Meier benchmark, suggesting that flexible parametric baselines are competitive for survival prediction.","The model gives comparable survival estimates to Cox PH and Kaplan-Meier at early time points, while adding closed-form shape and scale parameters that describe early- and late-risk subpopulations.","Because the model currently excludes covariates, it is positioned as a complement to Cox PH rather than a replacement: Cox contributes covariate interpretability, while EEPHND contributes a flexible closed-form baseline survival function."],"supporting_citations":[{"why":"Supplies the EEPD component's density, CDF, and inverse-transform sampling equations used inside the mixture.","marker":"[12]"},{"why":"Provides the mixture-construction reasoning and the moments and skewness/kurtosis framework applied to EEPHND.","marker":"[17]"},{"why":"Defines the concordance index used as the predictive-accuracy metric in the real-data comparison.","marker":"[22]"},{"why":"Supplies the Newton-Raphson numerical algorithm used for maximum likelihood estimation.","marker":"[32]"},{"why":"Provides the public lung-cancer survival dataset used for the real-data validation.","marker":"[33]"},{"why":"Documents the clinical-trial source of the lung-cancer survival data behind the benchmark.","marker":"[35]"}],"fun_headline_variants":["Mixture model beats Cox and Kaplan-Meier on lung cancer","New distribution hits 0.9997 concordance in survival","Closed-form mixture rivals nonparametric survival models","EEPHND outshines Cox PH in lung cancer concordance","Tractable mixture matches KM, beats Cox in survival fit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the plain product-of-densities likelihood, with no censoring weights, is valid for the right-censored lung cancer data, and that the concordance index computed in-sample from the fitted survival curve measures real predictive accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Mixture model beats Cox and Kaplan-Meier on lung cancer","New distribution hits 0.9997 concordance in survival","Closed-form mixture rivals nonparametric survival models","EEPHND outshines Cox PH in lung cancer concordance","Tractable mixture matches KM, beats Cox in survival fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00012,"raw_usage":{"total_tokens":1092,"prompt_tokens":951,"completion_tokens":141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":59}},"tokens_in":567,"tokens_out":141,"duration_ms":2759,"temperature":1.0,"reasoning_tokens":59,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:13:57.238137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit the model to the lung cancer data with a censoring-adjusted likelihood and evaluate the concordance index on held-out data, for example via time-dependent AUC with inverse-censoring weights; if the index falls toward the Cox level (near 0.6) or the fitted parameters shift sharply, the headline 0.9997 is an artifact of in-sample fitting.","supporting_citations":[{"cited_title":"Adeleke, A","cited_arxiv_id":null,"evidence_quote":"Supplies the EEPD component's density, CDF, and inverse-transform sampling equations used inside the mixture."},{"cited_title":"Adeleke, Beta-hyperhalfnormal distribution and its application, Basrah Journal of Sci- ence 38 (2) (2020) 131–156","cited_arxiv_id":null,"evidence_quote":"Provides the mixture-construction reasoning and the moments and skewness/kurtosis framework applied to EEPHND."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the concordance index used as the predictive-accuracy metric in the real-data comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Newton-Raphson numerical algorithm used for maximum likelihood estimation."},{"cited_title":"Davidson-Pilon, lifelines: survival analysis in python, Journal of Open Source Software 4 (40) (2019) 1317","cited_arxiv_id":null,"evidence_quote":"Provides the public lung-cancer survival dataset used for the real-data validation."},{"cited_title":"Jones, T","cited_arxiv_id":null,"evidence_quote":"Documents the clinical-trial source of the lung-cancer survival data behind the benchmark."}],"review_version":1}