{"id":"75bd0a14-961b-4c0a-8589-db1d1c49734b","arxiv_id":"2507.18773","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A cure-rate joint model with individual change points and an event-time upper bound on the change point estimates tumor burden trajectories and treatment effects including a stable patient subgroup.","lead":"This statistics paper builds a joint model of tumor burden and progression-free survival that separates patients into a stable, cure-like group and a change-point group whose tumor rebound is constrained by the event time. The model is tested on simulations and on a subset of a Phase 3 lung cancer trial, where it shows cemiplimab delays tumor regrowth.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 6's 'real data' application is contradicted by figure captions calling the same subset 'simulated'; the clinical conclusions hinge on resolving this.","rationale":"The reader's weakest_assumption was the PTMVN/log-normal distributional structure, which is a legitimate limitation and is acknowledged in the Discussion. However, within the paper's simulation design the estimator is at least self-consistent, and the log-normal AFT is explicitly described as illustrative. The more decisive issue is the unresolved real-versus-simulated status of the EMPOWER subset: it bears directly on the abstract and on Section 6, and unlike distributional misspecification it can be settled by a direct check of the code and a yes/no answer. I therefore treat provenance as the load-bearing concern rather than the distributional assumption, and I keep the reader's CONDITIONAL verdict. If the data are simulated, the application claims must be downgraded and Table 2 and Figure 5 should not be presented as evidence about cemiplimab versus chemotherapy; if the provenance is confirmed as real, the remaining conditions, such as reporting J and sensitivity to survival or random-effect misspecification, still justify conditional acceptance.","tokens_in":30590,"tokens_out":4710,"duration_ms":50866,"concrete_test":"Check the public GitHub repository (github.com/quyixiang/JoCuR) for the script that produces Section 6 results and determine whether it reads an actual EMPOWER data file or generates the 400-patient subset from the Section 5 simulation parameters. If the data are simulated, relabel Section 6 and Figure 5 as a synthetic illustration, remove or qualify the abstract's 'Phase 3 NSCLC trial' claim, and present Table 2 as simulation output; if the data are real, correct the Figure 1 and Figure 2 captions and specify data access. As a secondary verification, ask the authors to report the MCEM sample size J used in all simulations and in the application, since the E-step Monte Carlo error is otherwise unquantified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 states that the Kaplan-Meier curves and TB trajectories use 'a subset from the real data' (a random 400 of 710 EMPOWER subjects), while the captions of Figures 1 and 2 label the same display as a 'simulated EMPOWER study.' The text also says 'actual figures from the trial cannot be disclosed.' Section 6 then applies the model to 'the subset of the EMPOWER study mentioned in Section 2' and reports treatment-arm change points, slopes, and PD-L1 subgroup trends as clinical findings, and the abstract claims application to a Phase 3 NSCLC trial. These statements are internally inconsistent. If the subset is synthetic, the application section is a simulation presented as a real-data analysis, so the confidence intervals in Table 2 and the clinical conclusions do not support the stated external-validity claim; if it is real, the 'simulated' captions and the nondisclosure statement need correction. This ambiguity is load-bearing because the paper's central clinical utility claim depends directly on the provenance of the EMPOWER subset, and the manuscript as written does not resolve it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a cure-rate joint model for longitudinal tumor burden (TB) and time-to-event data in oncology trials. The population is partitioned into a stable ('cured') group, modeled with a linear mixed model, and a change-point group, modeled with a piecewise linear mixed model with an individual-specific change point that is constrained to occur before the log event time. A log-normal AFT model is used for the event time, and the random effects (including the change point) are assumed to follow a partially truncated multivariate normal distribution. Estimation is performed with a Monte Carlo EM algorithm, with closed-form M-step updates for most parameters and L-BFGS for the random-effect mean/covariance. The paper reports simulations comparing the full model to a change-point-only model [16] and an application to a subset of the EMPOWER-Lung 1 trial comparing cemiplimab to chemotherapy.","tokens_in":30804,"tokens_out":8309,"duration_ms":78553,"significance":"If the modeling framework is accepted, the paper makes a useful methodological contribution: it extends the reverse-conditional joint model of [16] to accommodate a stable/cured subgroup, provides a complete MCEM implementation with closed-form updates and publicly available code, and frames the target of inference through the estimand lens by focusing on marginal TB trajectories. The supplement contains detailed derivations and proofs for the E-step identities. However, the current evidence base is narrower than the paper's claims: the simulations only demonstrate internal consistency under the assumed parametric family, and the real-data application is clouded by conflicting statements about whether the EMPOWER subset is real or simulated. These issues currently limit the external validity claims made for the method.","major_comments":[{"comment":"Section 2 states that the Kaplan-Meier curves and TB trajectories use 'a subset from the real data' (a random 400 of 710 EMPOWER subjects) and that 'actual figures from the trial cannot be disclosed,' yet the captions of Figures 1 and 2 label the displays as a 'simulated EMPOWER study.' Section 6 then applies the model to 'the subset of the EMPOWER study mentioned in Section 2' and reports treatment-arm change points, slopes, and PD-L1 subgroup trends as clinical findings, with the abstract claiming application to a Phase 3 NSCLC trial. These statements are internally inconsistent. If the subset is synthetic, Section 6 is a simulation presented as a real-data analysis, so the confidence intervals in Table 2 and the clinical conclusions do not support the stated external-validity claim; if it is real, the 'simulated' captions and the nondisclosure statement need correction. This ambiguity is load-bearing because the paper's central clinical utility claim depends directly on the provenance of the EMPOWER subset.","section":"Section 2, Figures 1-2, Section 6"},{"comment":"The simulation section contains two related internal inconsistencies that undermine the reproducibility of the reported results. First, the symbol ωi is used both for a standard normal covariate in the AFT model ('We generate ωi with a standard normal distribution') and for the subject-specific change point in the PTMVN random effects in the same paragraph; the model in Section 3.2.2 uses w_i for the AFT covariate, so the simulation text should use w_i or another symbol. Second, the longitudinal outcome for the change-point group is written as 'yij ∼ N(xiβ1 + b0i + b1i(sij−ωi) + b2i(sij−ωi), σ2y)', which omits the indicator functions I(sij≤ωi) and I(sij>ωi) that define the piecewise model in Section 3.2.3; as written, both pre- and post-slopes apply at every visit time. Please correct these errors and clarify the data-generating algorithm.","section":"Section 5.1"},{"comment":"The simulation study generates data from the same parametric family the model assumes: log-normal AFT event times, PTMVN random effects, and piecewise linear longitudinal outcomes. The claim in Section 5.2 that the full model 'consistently maintains low bias and meets the 95% coverage target' is therefore a demonstration of internal consistency, not robustness. The comparison against the change-point-only model [16] is a comparison against a nested special case (π(s)=0) that is misspecified whenever the true π(s)>0, so the full model's superior performance is expected by construction. The paper should temper the 'robust' language and, ideally, add a misspecification scenario (e.g., Weibull event times or a different joint distribution for ω and survival) to assess the sensitivity of the change-point and trajectory estimates.","section":"Section 5.2"},{"comment":"The MCEM algorithm is central to estimation, but the manuscript does not specify the Monte Carlo sample size J used in the E-step (e.g., a fixed value, an increasing schedule with EM iterations, or a choice rule based on Monte Carlo error) or the convergence criterion for the EM iterations. Algorithm 1 takes J as an input, and the supplement describes sampling J draws, but the main text and supplement do not report how J was chosen for the simulations and the application. Without this information, the parameter estimates, the bootstrap confidence intervals in Table 2, and the trajectory CIs are not fully reproducible from the text alone.","section":"Section 4"}],"minor_comments":[{"comment":"The sentence 'ri = (ωi,bT i )T denotes the random effects for subject i who does not experience a change point' should read '...who experiences a change point' (or '...in the change-point group').","section":"Section 3.2.3"},{"comment":"The definition 'Bias = B−1|PB b=1(ˆθb−θ)|' appears to take the absolute value of the average bias rather than the average absolute bias; please clarify which quantity is reported, as this affects interpretation of the results.","section":"Section 5.2"},{"comment":"Figure 4 is very dense, with three panels and multiple scenarios; the 'X' marks indicating absent parameters are easy to overlook, and the figure would benefit from a legend or notes clarifying the meaning of the X marks.","section":"Figure 4"},{"comment":"In the subgroup analysis, the sample sizes (147, 120, 133) sum to 400, but it is not stated how these subgroups relate to the random 400-subject subset or whether any patients were excluded; please clarify the subgroup definitions.","section":"Section 6"},{"comment":"The notation E[y|s,θ] uses s for the scalar time point, while s_i is used for the vector of visit times in Section 3.1; a different symbol (e.g., s_new or t) would avoid ambiguity.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The provenance of the EMPOWER subset is the most consequential issue in this manuscript. The manuscript's internal inconsistency on this point is not a minor wording problem: if the 'real data' application is in fact a simulation, the paper's central claim of clinical utility is not supported by the evidence. I recommend asking the authors to state explicitly, in a prominent place, whether the EMPOWER subset is real patient data or synthetic data, and to adjust the abstract and Section 6 accordingly. Additionally, note that the baseline comparator model [16] is the authors' own prior work; while a nested comparison is acceptable, it would strengthen the paper to include an independent benchmark or at least acknowledge this in the discussion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece is the cure-rate mixture: a stable group modeled with a linear mixed model and a change-point group whose change point is bounded by the event time. That extends the authors' earlier change-point joint model and fixes a real misspecification when a stable subgroup exists. The derivations are careful, with closed-form M-step updates, proofs in the supplement, and public code. That is real value.\n\nThe soft spots, in proportion. The big one is the EMPOWER application. Section 2 says the motivating figures use \"a subset from the real data\" (400 of 710 subjects), while the captions of Figures 1 and 2 call the same display a \"simulated EMPOWER study,\" and the text says \"actual figures from the trial cannot be disclosed.\" Section 6 then applies the model to \"the subset of the EMPOWER study mentioned in Section 2\" and reports treatment-arm change points, slopes, and PD-L1 subgroup trends as clinical findings. Those statements cannot all be true. If the subset is synthetic, the application is a simulation presented as a real-data analysis, and the abstract's claim about a Phase 3 trial overstates what was done. If it is real, the captions and non-disclosure statement need correction. This is not a minor labeling issue; the paper's clinical utility claim depends on which it is.\n\nSecond, the simulation study validates the estimator only within the assumed parametric family—same AFT, same truncated normal structure, same longitudinal model. That shows internal consistency but not robustness. The authors acknowledge degraded parameter recovery when the pre/post slopes are similar, but they never test misspecification of the AFT or the PTMVN dependence. \"Outperforms\" should be scoped to the correct-model case.\n\nThird, the MCEM Monte Carlo size J is never stated, so the stochastic error in the E-step is unquantified. Minor, and easy to fix by reporting J and perhaps a sensitivity check.\n\nThe log-normal AFT is acknowledged as a limitation in the Discussion, and that is the right place for it. The deeper issue is that all the censoring and truncation information flows through that one distributional assumption.\n\nMy take: the methodological core deserves serious refereeing. The extension is sensible, the derivations are nontrivial, and the code helps reproducibility. The data-provenance problem is load-bearing and must be fixed. I would send this to referees, but with an explicit request to clarify what was real and what was simulated, and to temper the clinical language accordingly. I would not cite the clinical results until that ambiguity is resolved.","headline":"A careful extension of the authors' own change-point joint model, with a serious data-provenance ambiguity in the application that must be resolved before the clinical claims are taken at face value.","tokens_in":31342,"tokens_out":2348,"would_cite":false,"duration_ms":25872,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P10","62N01","62F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Separating the never-rebound group sharpens tumor-regrowth timing estimates.","keywords":["cure rate model","joint model","change point","tumor burden","time-to-event","MCEM","partially truncated multivariate normal","non-small cell lung cancer"],"falsifier":"Generate the same simulations with event times drawn from a Weibull distribution while the fitted model keeps its log-normal event-time assumption; if the marginal tumor-burden trajectory's 95% coverage falls substantially below nominal at late follow-up, the model's claimed accuracy depends on the assumed event-time distribution rather than on the cure-rate structure.","tokens_in":1590,"feed_emoji":"📈","tokens_out":3821,"duration_ms":104169,"temperature":0.7,"pith_summary":"This paper proposes a joint statistical model for two types of data in cancer trials: longitudinal tumor burden and time to progression or death. It separates patients into a stable group, whose tumor burden steadily declines, and a change-point group, whose tumor burden first shrinks and then later grows. For the change-point group, the event time is used to bound the individual change point rather than the other way around. Simulations show that when a stable subgroup exists, this mixture estimates the marginal tumor-burden trajectory with low bias and near-95% coverage, while a change-point-only model becomes increasingly biased as the stable fraction grows. Applied to a non-small-cell lung cancer trial, it estimates that cemiplimab delays the average tumor rebound and slows the post-rebound growth rate relative to chemotherapy.","feed_headline":"Tumor rebound timing sharpened by separating stable patients","feed_subtitle":"A cure-rate mixture estimates when tumors start regrowing, with better accuracy when some patients never rebound.","key_machinery":"The load-bearing object is the partially truncated multivariate normal distribution for the random effects of the change-point group, $f_{R|T}(r_i|t^*_i) \\propto \\varphi_4(r_i|\\mu_r, \\Sigma_r) \\, I(0 < \\omega_i \\le e^{t^*_i})$. It encodes the biological constraint that a patient's turnaround point $\\omega_i$ precedes the possibly latent progression time, and it is exactly the object the Monte Carlo E-step samples from when event times are censored. Around it, a log-normal accelerated failure time model supplies the event-time distribution, a piecewise linear mixed model supplies the longitudinal trajectory, and the marginal estimand $E[y|s,\\theta] = \\pi_{(s)}\\,\\mu_{(s)}(s) + (1-\\pi_{(s)})\\,\\mu_{(cp)}(s)$ turns the fitted mixture into a population-level tumor-burden curve. The M-step uses closed forms for the cure rate, regression coefficients, and variances, and L-BFGS on a Cholesky-reparameterized covariance for the random effects.","core_discovery":"The paper's central claim is that a cure-rate mixture of a stable group and a change-point group is the right structure for tumor-burden trajectories in oncology trials, with each change point constrained to precede the disease-progression event. Longitudinal tumor burden is modeled by a piecewise linear mixed model with random intercept, pre-slope, post-slope, and random change point; the change point is restricted to $(0, e^{t^*_i})$ through a partially truncated multivariate normal distribution, and the event time follows a log-normal accelerated failure time model. Estimation uses Monte Carlo EM with closed-form updates for most parameters and L-BFGS for the random-effects mean and covariance. In simulations with stable proportions $\\pi_{(s)} = 0.2$ and $0.4$, the full model keeps low bias and 95% coverage while a change-point-only model shows increasing bias and reduced coverage. In the application, the estimated overall change point is later for cemiplimab (0.556 years) than for chemotherapy (0.301 years), and the post-change slope is slower (0.091 vs 0.31).","pith_inferences":["A natural extension is to apply the same stable-versus-change-point mixture to other longitudinal biomarkers with a non-responding subset, such as PSA in prostate cancer, where the mixture would protect trajectory estimates from patients who never rebound.","The model's estimated probability of stable-group membership for each censored patient could be used as a patient-level durable-response classifier, a use the paper does not itself develop.","A Weibull or semiparametric event-time baseline would reveal how much of the trajectory accuracy is carried by the log-normal assumption, which the paper flags as a limitation."],"forward_implications":["When a stable subgroup exists, a change-point-only model misattributes stable patients to the change-point group, so its parameter bias grows and its coverage falls as $\\pi_{(s)}$ increases; the full mixture avoids this.","For censored patients, the time-to-event model supplies an upper bound on the latent change point, so rebound timing is estimated even when the rebound itself is never observed.","The marginal tumor-burden trajectory, the estimand of primary interest, is estimated with low bias and near-nominal coverage even when the pre- and post-slope difference is small, though individual parameters may be harder to identify.","In the EMPOWER application, cemiplimab is estimated to delay the overall change point (0.556 vs 0.301 years) and slow post-rebound growth (0.091 vs 0.31) relative to chemotherapy, with treatment differences separating after about 121 days."],"supporting_citations":[{"why":"Supplies the change-point-only model and the partially truncated multivariate normal distribution that the new model extends with a stable group.","marker":"[16]"},{"why":"Provides the EMPOWER-Lung 1 trial data and clinical context used in the application.","marker":"[17]"},{"why":"Supplies the estimands framing that motivates estimating marginal longitudinal trajectories rather than only parameters.","marker":"[18]"},{"why":"Supplies the regulatory estimands guidance that makes the marginal trajectory clinically meaningful.","marker":"[19]"},{"why":"Supplies the PD-L1 biology used to interpret the subgroup results.","marker":"[20]"}],"fun_headline_variants":["Cure-rate joint model sharpens tumor rebound timing","Separating stable patients improves change-point estimates","Change point before progression: joint model boosts accuracy","Mixture model tracks tumor rebound before progression event"],"cache_read_input_tokens":33536,"weakest_assumption_plain":"The model's load-bearing premise is that a progressing patient's true progression time is log-normal and that, given that time, the moment of tumor rebound follows the truncated normal shape the model assumes; if real tumors rebound on a different schedule, the censored patients are misweighted.","fun_headline_variants_meta":{"raw":{"variants":["Cure-rate joint model sharpens tumor rebound timing","Separating stable patients improves change-point estimates","Change point before progression: joint model boosts accuracy","Mixture model tracks tumor rebound before progression event"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":2017,"prompt_tokens":1006,"completion_tokens":1011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":952}},"tokens_in":622,"tokens_out":1011,"duration_ms":8540,"temperature":1.0,"reasoning_tokens":952,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:08:08.200843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate the same simulations with event times drawn from a Weibull distribution while the fitted model keeps its log-normal event-time assumption; if the marginal tumor-burden trajectory's 95% coverage falls substantially below nominal at late follow-up, the model's claimed accuracy depends on the assumed event-time distribution rather than on the cure-rate structure.","supporting_citations":[{"cited_title":"Jointly modeling time-to-event and longitudinal data with individual-specific change points: A case study in modeling tumor burden","cited_arxiv_id":null,"evidence_quote":"Supplies the change-point-only model and the partially truncated multivariate normal distribution that the new model extends with a stable group."},{"cited_title":"u m \\\"u s , Igor Bondarenko, Mustafa \\","cited_arxiv_id":null,"evidence_quote":"Provides the EMPOWER-Lung 1 trial data and clinical context used in the application."},{"cited_title":"The estimands framework: a primer on the ich e9 (r1) addendum","cited_arxiv_id":null,"evidence_quote":"Supplies the estimands framing that motivates estimating marginal longitudinal trajectories rather than only parameters."},{"cited_title":"Food and Drug Administration","cited_arxiv_id":null,"evidence_quote":"Supplies the regulatory estimands guidance that makes the marginal trajectory clinically meaningful."},{"cited_title":"Regulation and function of the pd-l1 checkpoint","cited_arxiv_id":null,"evidence_quote":"Supplies the PD-L1 biology used to interpret the subgroup results."}],"review_version":2}