{"id":"f698f8fc-9626-4af4-b2fd-f026288f5998","arxiv_id":"2501.16910","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":22,"one_line_summary":"A bilateral hidden Markov model trained on 833 OPSCC patients estimates personalized risk of occult contralateral lymph node disease and proposes smaller elective radiation volumes for selected patients.","lead":"The authors extend a hidden Markov model of lymph node spread in oropharyngeal cancer to predict the risk of occult disease on the opposite side of the neck. The model suggests many patients could receive less radiation to the contralateral neck, based on whether the tumor crosses the midline, T-stage, and ipsilateral involvement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"For advanced lateralized tumors with ipsilateral LNL II+III involvement, the paper's own training data (12/54, 22.2%) imply ~6.7% occult contralateral-II risk after a clinically negative workup, above the 5% threshold; yet §8.2 recommends unilateral RT 'regardless' of ipsilateral involvement.","rationale":"I focused on the central safety claim rather than the structural assumption the reader flagged. The no-direct-drainage assumption (Eq. 9) is a legitimate concern, but it is a modeling assumption that can be tested by model comparison and is partially mitigated by the model's ability to reproduce ipsi-contralateral correlations through shared time. The issue I identify is more immediate and more damaging: the paper's own reported data and threshold appear to contradict its blanket recommendation. The calculation is simple and uses only numbers stated in the manuscript: 12/54 from Section 2.3.4, sensitivity 81% and specificity 76% from Section 7, and the 5% threshold from Section 8.2. The resulting posterior risk of 6.7% is not a subtle bias; it is a direct violation of the stated criterion if the model is well calibrated, or a failure of the model's in-sample fit if it is not. The reader's verdict of CONDITIONAL remains appropriate because this concern is checkable with the released code and, if confirmed, would require narrowing the recommendation to exclude high-risk lateralized subgroups. I do not think it warrants outright rejection, since the modeling framework and data release are valuable and the clinical conclusion can be revised.","tokens_in":18555,"tokens_out":15548,"duration_ms":148320,"concrete_test":"Run the released code (rmnldwg/bilateral-paper) for the exact scenario: advanced T-category, epsilon=False (lateralized), ipsilateral LNL II and III clinically involved, all contralateral LNLs clinically negative, and compute the posterior P(contralateral LNL II involved) using Eq. 16 with the MCMC posterior samples. Compare this value to 5% and to the 6.7% derived from the 12/54 data point. Also print the model's predicted prevalence for this scenario and overlay it on the beta posterior from n=54, k=12 in Fig. 9 bottom-left red; if the model's posterior exceeds 5% the 'regardless' recommendation is internally inconsistent, and if it is below 5% the model visibly underfits this safety-critical training cell.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is not the graph structure but the consistency between the paper's own data and its central recommendation. Section 8.2 states that for lateralized tumors with no contralateral clinical involvement, unilateral radiotherapy is sufficient 'regardless of T-category or ipsilateral involvement,' based on a 5% acceptable-risk threshold. Section 2.3.4 reports that, among patients with lateralized, advanced T-category tumors and ipsilateral LNLs II+III involved, 12/54 (22.2%) had contralateral LNL II involvement. For a patient in this exact subgroup whose contralateral LNL II is clinically negative, applying the paper's own imaging characteristics from Section 7 (sensitivity 81%, specificity 76%) via Bayes' rule gives P(X=1 | Z=0) = 0.19 × 0.222 / (0.19 × 0.222 + 0.76 × 0.778) ≈ 6.7%, above the 5% threshold. Thus either the model's posterior for this subgroup is >5%, in which case the blanket de-escalation recommendation contradicts the model's own criterion, or the model's posterior is <5%, in which case the model fails to reproduce the 22.2% prevalence it was trained on (the scenario in Fig. 9, bottom-left red panel), undermining the claim of accurate risk estimates. Section 8.3's limitations do not address this safety-relevant subgroup.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends a previously published hidden Markov model for ipsilateral lymphatic spread in oropharyngeal squamous cell carcinoma (OPSCC) to the contralateral neck. The bilateral model represents lymph node levels (LNLs) I, II, III, IV, V, and VII on both sides as binary hidden variables, evolves them synchronously over time, and couples the two sides only through the shared time to diagnosis and the primary tumor's midline extension. Parameters are inferred via MCMC from a pooled multi-institutional dataset of 833 patients. The authors report that the model reproduces observed contralateral involvement patterns, estimates risks of occult contralateral disease, and, based on a 5% risk threshold, recommends unilateral elective irradiation for lateralized tumors and limitation to LNL II for midline-crossing tumors with a clinically negative contralateral neck.","tokens_in":18944,"tokens_out":12601,"duration_ms":107504,"significance":"If the risk estimates are reliable, the model could support personalized CTV-N de-escalation in a substantial patient group, with meaningful toxicity reduction. The paper's strengths include a compact and interpretable parameterization, a multi-institutional dataset, a reproducible implementation with publicly available code and data, and explicit MCMC convergence diagnostics. The clinical significance is, however, conditional on the model's ability to generalize beyond the training cohort and on the safety of the proposed risk threshold for the highest-risk subgroups; the current manuscript does not yet establish those conditions.","major_comments":[{"comment":"The model evaluation in §6.2 is entirely in-sample: the predicted prevalences in Figures 7–9 are compared with the observed prevalences in the same 833 patients used for training (§5.1). Because the model parameters are fit to these data, the agreement shown in those figures is expected and does not by itself demonstrate that the model will provide accurate risk estimates for new patients. Given that the paper makes a clinical recommendation (unilateral CTV-N exclusion for lateralized tumors), the authors should provide at least an internal validation, for example by cross-validation or by training on three of the four institutions in Table 1 and testing on the fourth. Without such validation, the claim that the model 'accurately and precisely describes observed patterns' is overstated.","section":"§5.1, §6.2"},{"comment":"The blanket recommendation in §8.2 that for lateralized tumors with no contralateral clinical involvement 'unilateral radiotherapy is sufficient, regardless of T-category or ipsilateral involvement' is not reconciled with the paper's own data. Section 2.3.4 reports a 22.2% prevalence (12 of 54 patients) of contralateral level II involvement for advanced T-category lateralized tumors with ipsilateral levels II and III involved. For a patient in this subgroup with a clinically negative contralateral level II, the imaging characteristics assumed in §7 (sensitivity 81%, specificity 76%) yield a posterior occult-disease risk of about 6.7%, above the 5% threshold used throughout the paper. Conditioning on additional contralateral levels being clinically negative could lower this posterior, but the paper does not report the model's predicted risk for this exact subgroup in §7.1 or Figure 10. Since this is the subgroup for which the 'regardless' recommendation is most consequential, the authors should either report the model's posterior for this scenario and demonstrate that it falls below 5%, or qualify the recommendation. As written, the central de-escalation claim is not directly supported by the presented risk estimates.","section":"§8.2 with §2.3.4 and §7.1"},{"comment":"The training procedure treats the consensus involvement states as the true hidden states X (§5.1), while the posterior risk predictions in §7 apply an observation model with sensitivity/specificity to the same LNLs. For non-surgical patients the consensus is itself derived from imaging using the same sensitivity/specificity values (§B, Table 3), so the training targets already incorporate diagnostic uncertainty that is later re-applied as if it were the clinical observation process. The paper acknowledges this approximation in §8.3.2, but it does not assess how sensitive the occult-risk estimates are to it. A sensitivity analysis—for example, training the model only on pathologically confirmed cases and validating on clinically diagnosed cases, or treating the consensus as a latent variable—would strengthen confidence in the reported risk estimates.","section":"§5.1, §B, §8.3.2"}],"minor_comments":[{"comment":"There are several typos: 'diagnositc' in Section 3 after Eq. (3), 'enumator' in §5.3, 'matrx' in §4.3, and 'probabiltiy' in §3.1; these should be corrected.","section":"§3, §5.3, §4.3"},{"comment":"The y-axis is omitted in Figures 7–9; the authors state that the numerical value is not intuitively interpretable, but this makes it difficult to assess whether the predicted histograms match the width of the observed beta posteriors. Consider adding a labeled axis or a scale.","section":"Figures 7–9"},{"comment":"In the recursive formula for P(Xc, ϵ=True|τ+1), the variable τ is used without explicit definition; please clarify that τ indexes time steps from 0 to tmax.","section":"§4.2"},{"comment":"The sensitivity and specificity values are introduced for 'imaging' but the numbers correspond to CT in Table 3; please specify the modality (e.g., CT) to avoid ambiguity for readers.","section":"§7"}],"recommendation":"major_revision","confidential_remarks":"The authors are established in this niche (prior ipsilateral model, LyProX platform), and the manuscript is a natural progression. The main concerns are that the clinical recommendation is based on in-sample fitting and that the 'regardless' recommendation for lateralized tumors is not backed by an explicit risk estimate for the highest-risk lateralized subgroup. I would not reject the paper: the model is plausible and the issues are addressable with additional analyses and a more cautious wording. The statement that the model is already guiding a clinical trial strengthens the need for the authors to close these gaps before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core contribution is real: a bilateral extension of the ipsilateral HMM with a midline-extension latent variable, trained on 833 OPSCC patients with MCMC, and fully reproducible code/data. The model is interpretable, the parameter estimates are sensible, and the paper is honest about several limitations (consensus-as-truth approximation, midline-extension prevalence mismatch by T-category). For a modeling paper, this is well above average.\n\nBut there's a load-bearing internal inconsistency that the authors don't address. Section 8.2 recommends unilateral RT for lateralized tumors with no contralateral clinical involvement, 'regardless of T-category or ipsilateral involvement,' based on a 5% acceptable risk threshold. Section 2.3.4 reports that among advanced T-category lateralized tumors with ipsilateral LNLs II+III involved, 12/54 (22.2%) had contralateral LNL II involvement. Using the paper's own imaging sensitivity (81%) and specificity (76%) from Section 7, a clinically negative contralateral II in that subgroup has posterior occult risk ~6.7% — above 5%. So either the model's posterior for this subgroup is >5%, in which case the blanket recommendation contradicts the model's own criterion, or the model's posterior is <5%, in which case the model fails to reproduce the prevalence it was trained on. The limitations section (8.3) doesn't mention this safety-relevant subgroup.\n\nThe other soft spots are milder. The evaluation is entirely in-sample; there is no held-out test set or cross-validation. The 'consensus' as truth is an approximation, acknowledged but potentially biasing both training and evaluation. The midline-extension prevalence mismatch is acknowledged and partly mitigated by conditioning, but it still affects the joint distribution. These are not fatal, but they reinforce that the clinical safety claims need external validation before they guide de-escalation.\n\nOverall: this deserves peer review. The modeling is serious, the data/code are reproducible, and the clinical question is important. But the authors need to fix the threshold inconsistency — either refine the recommendation for the high-risk subgroup or show that the model's posterior for that subgroup is actually below 5% and explain why it deviates from the 22.2% prevalence. I'd send it to a journal with a request for that revision and for out-of-sample validation.","headline":"A solid, interpretable bilateral HMM with open code and a real internal contradiction between its de-escalation rule and its own data.","tokens_in":19482,"tokens_out":3089,"would_cite":true,"duration_ms":26925,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A probabilistic model maps which throat cancer patients can safely skip contralateral neck radiation.","keywords":["oropharyngeal squamous cell carcinoma","lymph node levels","occult metastasis","contralateral neck","elective clinical target volume","hidden Markov model","midline extension","CTV-N de-escalation"],"falsifier":"Measure occult contralateral disease in a prospective cohort of lateralized T1-T2 oropharyngeal tumors with clinically negative contralateral necks who undergo elective contralateral neck dissection, and check whether contralateral level II involvement significantly exceeds the model's predicted risk; a proportion above roughly 5% would invalidate the de-escalation recommendation.","tokens_in":18315,"feed_emoji":"🎗️","tokens_out":5641,"duration_ms":50722,"temperature":0.7,"pith_summary":"This paper tries to establish that the risk of undetected ('occult') cancer spread to the lymph nodes on the side of the neck opposite an oropharyngeal tumor can be predicted per patient from clinical facts known before treatment: tumor position relative to the midline, T-category, and which lymph node levels are already involved on the same side. If this is right, many patients with lateralized tumors and a clinically negative contralateral neck can safely have the contralateral side left out of the elective radiation target volume, and patients with midline-crossing tumors can have that volume reduced to lymph node level II. The argument is carried by a bilateral hidden Markov model of lymphatic progression trained on 833 patients across four institutions, which reproduces observed bilateral involvement patterns with a compact set of interpretable parameters. The model identifies midline extension as the dominant risk factor for contralateral spread, with advanced T-stage and heavier ipsilateral involvement adding further risk.","feed_headline":"Lateralized throat tumors may skip contralateral neck radiation","feed_subtitle":"A hidden Markov model of lymphatic spread says midline-crossing tumors need only level II irradiation.","key_machinery":"The engine is a hidden Markov model over the six lymph node levels I, II, III, IV, V, and VII on each side, where the hidden state is healthy or involved and observed through noisy imaging with fixed sensitivity and specificity. The bilateral extension keeps the ipsilateral transition graph, adds a contralateral graph with shared inter-level spread rates, and couples the sides through a time-prior that differs for early versus advanced T-category; the conditional independence $P(X^i, X^c \\mid t) = P(X^i \\mid t)\\,P(X^c \\mid t)$ encodes the assumption that contralateral disease spreads only from the primary tumor. Midline extension is a binary random variable $\\epsilon$ that switches the contralateral tumor-spread rates from $b^{c,\\epsilon=\\mathrm{False}}_v$ to a linear mix $\\alpha\\,b^i_v + (1-\\alpha)\\,b^{c,\\epsilon=\\mathrm{False}}_v$, with $\\alpha \\approx 0.34$ in the fitted model. Parameters are learned by Markov chain Monte Carlo from 833 patients, and posterior risks for occult disease in any level are computed by marginalizing the joint distribution over hidden states.","core_discovery":"The central claim is that bilateral lymphatic progression in oropharyngeal squamous cell carcinoma is explained by a single shared time axis: the two sides of the neck evolve independently, each governed by the same spread dynamics between lymph node levels, and are coupled only through the diagnosis time and through tumor spread rates that differ ipsilaterally versus contralaterally. Under this construction, contralateral involvement is driven primarily by whether the primary tumor crosses the midline (encoded as a random switch with probability $p_\\epsilon$ per time step), with advanced T-stage and ipsilateral burden acting as secondary risk factors through later diagnosis times. For a lateralized tumor with a clinically negative contralateral neck, the model puts the occult contralateral level II risk near or below a 5% threshold across T-categories, and predicts that contralateral level III involvement is unlikely without level II involvement and level IV involvement is rare without level III, leading to the paper's volume de-escalation recommendations.","pith_inferences":["If confirmed prospectively, the same two-sided conditional-independence structure could be transferred to other head and neck subsites such as the oral cavity, hypopharynx, and larynx, where bilateral elective irradiation is standard.","The 5% occult-risk threshold is a clinical policy choice rather than a model output; a reader who prefers a stricter threshold would shrink or expand the recommended volumes accordingly.","A direct test of the no-cross-midline-drainage assumption could come from lymphoscintigraphy, or from comparing model predictions to pathological findings in elective contralateral neck dissections, which would reveal whether rerouted drainage after bulky ipsilateral disease is clinically significant.","Because the model outputs continuous per-level risk estimates, it could be connected to treatment planning to quantify expected reductions in xerostomia and dysphagia from unilateral or level-II-only irradiation."],"forward_implications":["For lateralized tumors with no clinical contralateral involvement, the contralateral neck can be excluded from the elective CTV-N regardless of T-category or ipsilateral involvement, according to the 5% occult-risk threshold.","For tumors crossing the midline with a clinically negative contralateral neck, elective contralateral irradiation can be limited to lymph node level II.","Contralateral level III should be irradiated only when level II is clinically involved, and contralateral level IV only when level III involvement is confirmed.","Contralateral levels I, V, and VII generally do not require elective irradiation unless they are clinically involved.","The paper notes that the model's predictions are already guiding a clinical trial on volume de-escalation."],"supporting_citations":[{"why":"Supplies the ipsilateral hidden Markov model and the directed acyclic graph of lymph node level spread that this work extends to the contralateral side.","marker":"[16]"},{"why":"Earlier formulations of the unilateral lymphatic progression model and its parameter values, which the bilateral model reproduces.","marker":"[9, 10]"},{"why":"Compile the multi-institutional patient dataset on per-level lymph node involvement used to train and evaluate the model.","marker":"[12, 13]"},{"why":"Provide the sensitivity and specificity values for CT, PET, MRI, FNA, and pathology used to build the consensus diagnosis and observation model.","marker":"[14, 15]"},{"why":"Introduced the Bayesian network approach to personalized elective CTV definition that the hidden Markov formulation builds on.","marker":"[11]"},{"why":"Registers the clinical trial in which the model's volume de-escalation predictions are currently being tested.","marker":"[22]"}],"fun_headline_variants":["Skip contralateral neck radiation for lateralized throat tumors","Contralateral neck radiation may be unnecessary for lateralized throat tumors","Midline-crossing throat tumors may need only level II radiation","Contralateral neck radiation is optional for lateralized throat tumors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes the two sides of the neck never drain into each other and that contralateral spread happens only through the primary tumor; if lymphatic rerouting or cross-midline connections occur, the predicted contralateral risks are too low.","fun_headline_variants_meta":{"raw":{"variants":["Skip contralateral neck radiation for lateralized throat tumors","Contralateral neck radiation may be unnecessary for lateralized throat tumors","Midline-crossing throat tumors may need only level II radiation","Contralateral neck radiation is optional for lateralized throat tumors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001217,"raw_usage":{"total_tokens":5060,"prompt_tokens":1053,"completion_tokens":4007,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":3946}},"tokens_in":669,"tokens_out":4007,"duration_ms":25006,"temperature":1.0,"reasoning_tokens":3946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:42:38.306684+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure occult contralateral disease in a prospective cohort of lateralized T1-T2 oropharyngeal tumors with clinically negative contralateral necks who undergo elective contralateral neck dissection, and check whether contralateral level II involvement significantly exceeds the model's predicted risk; a proportion above roughly 5% would invalidate the de-escalation recommendation.","supporting_citations":[{"cited_title":"Scientific Reports14(1), 15750 (2024) https://doi.org/10.1038/s41598-024-66012-1","cited_arxiv_id":null,"evidence_quote":"Supplies the ipsilateral hidden Markov model and the directed acyclic graph of lymph node level spread that this work extends to the contralateral side."},{"cited_title":"Physics in Medicine & Biology64(16), 165003 (2019) https://doi.org/10.1088/1361-6560/ab2a18","cited_arxiv_id":null,"evidence_quote":"Introduced the Bayesian network approach to personalized elective CTV definition that the hidden Markov formulation builds on."},{"cited_title":"Clinical Trial Registration NCT06563362, clinicaltrials.gov (August 2024)","cited_arxiv_id":null,"evidence_quote":"Registers the clinical trial in which the model's volume de-escalation predictions are currently being tested."}],"review_version":1}