{"id":"f30d860c-c571-45c0-b3dd-4eb090bcdbcf","arxiv_id":"2505.05700","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A shape-constrained Bayesian spline model fits S-shaped Alzheimer's biomarker trajectories and estimates their timing and ordering from the BIOCARD study.","lead":"This paper introduces a Bayesian statistical model that forces disease biomarker curves to follow a smooth S-shape, then applies it to Alzheimer's data. The model estimates when each biomarker starts changing and which ones change first, while capturing differences between early and late disease stages.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The intercept-only heterogeneity in Eq. (3.6) cannot represent subject-specific disease onsets or progression rates; if individual trajectories are S-shaped in disease stage rather than age, the estimated BIOCARD milestone ordering may be a mixture artifact rather than a validated temporal ordering.","rationale":"The reader identified the shared inflection index m* in Section 3.2 as the weakest assumption. That is a real constraint, but its main effect is within biomarker categories, whereas the abstract's headline ordering is across the three categories and each category has its own m*. The more load-bearing assumption is the absence of subject-level onset and rate variation in Eq. (3.6). This assumption is directly connected to the applied claim: the milestones used to infer temporal ordering are functionals of a single population curve, and if individuals vary in onset and rate, that curve is a mixture rather than a biological trajectory. The paper itself flags this limitation in Section 6, so it is acknowledged rather than hidden. A simulation test that generates from a stage-shifted S-curve model and fits the intercept-only model would settle whether the concern matters in practice. If the ordering survives, the applied claim is strengthened; if not, the temporal-ordering conclusion should be downgraded or the model extended. Since the methodological contribution, shape-constrained splines with Bayesian fitting, is independent and the simulation study is honest about boundary sensitivity, I would not reject the paper. I would keep the conditional verdict and require the additional simulation or a sensitivity analysis before the BIOCARD ordering claim is treated as established.","tokens_in":17958,"tokens_out":8182,"duration_ms":93924,"concrete_test":"Simulate from a generative model that matches the Section 4 design but adds subject-specific onset ages tau_i and rates r_i: y_ijk = f_k((t_ij - tau_i)/r_i) + x_i^T beta_k + epsilon_ijk, with f_k chosen as S-shaped curves whose true ordering is known across the CSF/MRI/cognitive categories. Fit the paper's actual model (Eq. 3.6) with the same priors, M=24 basis knots, and the [0,120] boundary. If the posterior ordering of the estimated inflection points or 50% thresholds across categories differs from the generating truth, or if the between-replicate credible intervals for the milestone ages straddle the ordering, then the Section 5.2 temporal-ordering conclusion is not identifiable from the intercept-only model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 (Eq. 3.6) models between-subject variation solely through additive intercepts omega_ik. There is no subject-specific age shift or rate parameter. Section 6 explicitly defers 'modeling heterogeneity across individuals not only in their baseline biomarker values but also in their disease onsets and rates of disease progressions' to future work. For late-onset Alzheimer's disease, however, this is not an optional refinement: biomarker trajectories are widely thought to be S-shaped functions of time since disease onset, with onset ages and progression rates varying across individuals. When such subjects are pooled by chronological age, the population-level curve estimated by Eq. (3.6) is a mixture of stage-shifted S-curves. A monotone S-shaped f_k(t) can still fit this mixture, but its unique inflection point and 50% threshold, the milestones used for the temporal-ordering conclusions in Section 5.2 and standardized in Eq. (3.7), need not equal any individual's milestones and can be systematically shifted by the distribution of onset ages and by the selection of cognitively normal participants at enrollment. The BIOCARD claim that CSF biomarkers precede MRI and cognitive biomarkers is therefore not securely established by this model; it is at least partly an artifact of the additive-intercept assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Bayesian shape-constrained spline regression model for Alzheimer's disease biomarker trajectories. The method builds on monotone regression splines and adds two constraints: a unique inflection point (imposed through unimodality of the spline coefficients) and a vanishing derivative at the boundaries. The model is fitted with a Gibbs sampler that handles the constrained parameter space. The authors evaluate the approach in a simulation study against a parametric logistic model and an unconstrained monotone spline model, and then apply it to BIOCARD data with eleven biomarkers in three categories. The main applied claim is that the estimated curves reveal a temporal ordering consistent with current hypotheses, with CSF biomarkers changing before MRI biomarkers, which change before cognitive biomarkers.","tokens_in":18286,"tokens_out":5629,"duration_ms":68813,"significance":"If the method is sound, it provides a useful semi-parametric tool for modeling S-shaped biomarker trajectories with interpretable milestones, and it could generalize to other neurodegenerative diseases. The paper's strengths include a novel combination of shape constraints, a full Bayesian fitting procedure with careful attention to constrained sampling, a simulation study that benchmarks against both parametric and overly flexible alternatives, and publicly available code. The real-data analysis demonstrates the method's ability to produce clinically coherent ordering, and the simulation results support the advantage of the S-shaped model over unconstrained monotone splines, particularly for asymmetric true curves. However, the significance is tempered by three concerns: the credible intervals show poor frequentist coverage, the subject-level heterogeneity is limited to additive intercepts, and the within-category shared inflection index is an untested assumption with potential impact on the real-data conclusions.","major_comments":[{"comment":"The applied temporal-ordering conclusion is not securely established because the model accounts for subject heterogeneity only through additive intercepts omega_ik. The paper itself (Section 6) defers modeling of between-subject variation in disease onset and progression rate to future work. If individual biomarker trajectories are S-shaped in time since onset rather than in age, pooling subjects by chronological age with only intercept shifts can produce a population-level curve whose inflection point and 50% threshold are mixture artifacts, not biologically meaningful milestones. The simulation study (Section 4.1) generates data from model (3.6) and therefore cannot validate the milestone estimates under realistic onset heterogeneity. I recommend either adding a simulation that includes subject-specific onset/rate shifts or substantially tempering the claim that the BIOCARD ordering is a validated scientific finding.","section":"Section 5.2, Eq. (3.6), Section 6"},{"comment":"The S-shaped model's 95% credible intervals severely under-cover the true curve: 62-69% under flogit and 55-69% under fasym, with the worst values at the recommended wider boundary [0,120]. The intervals also over-cover the milestones (100% for flogit). This contradicts the stated advantage in Section 2.5 of Bayesian uncertainty quantification and is not explained or remedied in the paper. I would like the authors to investigate the source of the miscalibration (for example, the informative prior on the penalty parameters or the hard boundary constraints) and to report either adjusted intervals, a sensitivity analysis, or a clear discussion of why the under-coverage is acceptable.","section":"Table 1, Section 4.2"},{"comment":"All biomarkers within the same category are forced to share the same inflection index m* in Equation (2.3). The simulation study uses a single biomarker and therefore never exercises this group-level constraint. The real-data curves and their milestones can be directly affected by this assumption, yet no sensitivity analysis is provided (for example, allowing biomarker-specific m* or different category groupings). The paper should at least discuss the implications of this constraint and ideally provide a sensitivity check, since the within-category ordering claim is not the focus but the constraint may influence the estimated category-level curves.","section":"Section 3.2, Section 4.1"}],"minor_comments":[{"comment":"The manuscript uses 'ad' interchangeably with 'Alzheimer's disease' and 'biocard' for 'BIOCARD'; please normalize these to standard capitalization for readability.","section":"Throughout"},{"comment":"The simulation description says 'the baseline biomarker value omega_i' but the model in Eq. (3.6) defines omega_ik; since the simulation has a single biomarker, please clarify the notation and state that the subscript k is dropped.","section":"Section 4.1"},{"comment":"The prior written for (sigma_s^2, sigma_v^2) appears to place an exponential prior on the variance parameters themselves rather than a half-normal prior on the standard deviations as described in Section 3.2; please clarify the exact prior specification and its relation to the text.","section":"Supplement S2"},{"comment":"The legend lists six MRI biomarkers but the plot and text refer to five; also the vertical bars for inflection points are difficult to distinguish across 11 overlapping curves. Consider enlarging the plot or using a separate panel per category.","section":"Figure 2"},{"comment":"The learning-effect adjustment is estimated in a two-stage procedure, but the uncertainty in the learning-effect estimates is not propagated into the main analysis; this is worth mentioning as a limitation.","section":"Supplement S4"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a solid methodological contribution, but the applied conclusion in the abstract is stronger than what the model can currently support. The under-coverage of the credible intervals is a substantive statistical issue that should be addressed before publication. The comparison against the flexible monotone spline is informative, but the real-data analysis would benefit from a more cautious framing given the additive-intercept limitation. I believe the paper is within the scope of the journal and can be revised to address these concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a worthwhile paper. The methodological core is real, the simulation is honestly reported, and the authors don't oversell the clinical implications. The abstract does oversell the applied conclusion, though. The BIOCARD ordering is less secure than it looks.\n\nWhat's actually new: combining monotone splines, a unimodal derivative constraint, and a vanishing-boundary penalty into one Bayesian hierarchical model, with inference via truncated Gaussian sampling. The pieces exist separately, but the combination is useful and the implementation looks careful. I especially appreciate that the simulation compares against a correctly specified logistic model, which tilts things in the parametric model's favor, and still the S-shaped model holds up well. Under asymmetric truth it clearly wins. The under-coverage in some simulation cells (62–69% for the curve under flogit) is real but not disqualifying, and the authors discuss it.\n\nSoft spots, in rough order of importance. First, the shared inflection index for biomarkers within a category is a strong prior constraint. The within-category ordering of milestones is therefore partially built into the model, not learned freely, and there's no sensitivity analysis for that choice. Second, the between-subject heterogeneity is just an additive intercept. The stress-test note is on target: if individual trajectories are S-shaped in time-since-onset rather than age, then the population-level curve is a mixture of stage-shifted S-curves, and its inflection point and 50% threshold are not any individual's milestones. The authors explicitly defer onset and rate heterogeneity to future work in the discussion, so they know, but that means the BIOCARD temporal-ordering claim should be read as conditional on a fairly restrictive heterogeneity model. Third, there's no real-data baseline comparison—they don't fit the logistic or flexible spline to BIOCARD, so we don't see whether the S-shaped model changes the applied conclusions. Fourth, missing convergence diagnostics. Minor, but for an applied Bayesian paper I'd want trace plots or R-hats.\n\nWho is this for? Biostatisticians working on AD biomarker trajectories and anyone interested in practical shape-constrained splines. The methodology is sound and transferable to other S-shaped progressions. I would engage with it and I'd send it out. It deserves a serious referee, and a good revision will address the shared-inflection sensitivity and the interpretation of population-level milestones under heterogeneous onsets.","headline":"Solid and useful shape-constrained spline method, but the applied temporal-ordering claim is weaker than the abstract implies because heterogeneity is intercept-only and the shared inflection index is imposed a priori.","tokens_in":18767,"tokens_out":2296,"would_cite":true,"duration_ms":27781,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62F15","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"S-shaped splines recover Alzheimer's biomarker progression order.","keywords":["shape-constrained regression","monotone spline","S-shaped curves","Bayesian hierarchical model","Alzheimer's disease biomarkers","temporal ordering","BIOCARD","inflection point"],"falsifier":"Refit the BIOCARD analysis with each biomarker allowed its own inflection index instead of one shared per category; if the posterior credible intervals for inflection ages within any category separate by several years, the shared-index constraint that shapes the paper's within-category ordering would be contradicted.","tokens_in":17788,"feed_emoji":"🧠","tokens_out":6935,"duration_ms":72665,"temperature":0.7,"pith_summary":"This paper argues that Alzheimer's disease biomarker trajectories are best quantified as S-shaped curves of age: monotonically increasing, with one turning point where the change is fastest, and flattening at both early and late ages. The authors build a Bayesian spline regression that enforces these three shape constraints directly on the spline coefficients, and show in simulations that when the true progression is asymmetric the constrained model estimates the curve and its milestone ages more accurately than either a logistic parametric curve or an unconstrained monotone spline. Applied to the BIOCARD longitudinal study, the model estimates that cerebrospinal fluid biomarkers turn first, MRI measures of brain structure second, and cognitive tests last, matching the scientific hypothesis known as the Jack model. The value of the paper is a middle path between rigid parametric curves and overly flexible nonparametric ones, with interpretable milestones and quantified uncertainty.","feed_headline":"S-shaped splines order Alzheimer's biomarkers in time","feed_subtitle":"Bayesian curves with one turning point place CSF before MRI before cognitive decline in BIOCARD data.","key_machinery":"The central object is an S-shaped regression spline $f(t)=\\sum_{m=1}^{M} \\gamma_m I_m(t)$ built from integrated quadratic B-spline bases $I_m(t)$. Three coefficient constraints carry the argument: non-negative coefficients $\\gamma_m$ ensure monotone increase; an inflection index $m^*$ with $\\gamma_1 \\le \\cdots \\le \\gamma_{m^*} \\ge \\cdots \\ge \\gamma_M$ makes the derivative unimodal, giving exactly one inflection point; and boundary coefficients $\\gamma_1, \\gamma_2, \\gamma_{M-1}, \\gamma_M$ are set to zero with a Planck-taper window penalty, sending the derivative to zero at both ends. The Bayesian hierarchical version adds subject-level random intercepts and estimates a shared inflection index for each group of biomarkers, with posterior inference via a Gibbs sampler using truncated multivariate Gaussian draws and multivariate normal probability computations.","core_discovery":"The central claim is that adding two scientifically motivated constraints to monotone regression splines—vanishing derivative at the boundaries and a unique inflection point—turns the spline into an interpretable S-shaped curve without sacrificing the flexibility needed to capture asymmetric progression. The paper develops this shape-constrained spline in a Bayesian hierarchical model with subject-level random intercepts, and fits it to eleven BIOCARD biomarkers. Its key finding is the estimated temporal ordering: CSF biomarkers reach their 50% thresholds and inflection points earliest, followed by MRI biomarkers, with cognitive tests last; this ordering is consistent with the Jack hypothesis and overlaps sensibly with the age distributions of symptom onset and dementia diagnosis. In simulation, the S-shaped model recovers the true curve, inflection point, and 50% threshold with lower error than the logistic model when the true curve is asymmetric, while the unconstrained monotone spline performs poorly because it cannot represent early and late plateaus and over-extrapolates outside the observed age range.","pith_inferences":["The same S-shaped spline could be coupled with a latent disease-progression time that shifts and stretches each individual's curve, so that heterogeneity in onset and progression rate is modeled rather than only baseline levels; this would address a limitation the paper leaves open.","A direct test of the plateau assumption would come from longer follow-up: if biomarkers keep declining slowly after the plateau age, the zero-boundary-derivative constraint would bias late-life estimates, and the bias could be detected by out-of-sample prediction on late ages.","The method's transfer to Parkinson's and Huntington's disease is plausible, but those disorders may have different asymmetry directions and plateau structures, so the Planck-taper window and boundary choices would need recalibration.","A natural robustness check is to vary the spline knot boundaries—the paper shows wider boundaries improve coverage under asymmetric truth—and to report whether the estimated biomarker ordering persists across those choices."],"forward_implications":["Posterior samples give full uncertainty quantification for milestone ages (50% threshold and inflection point), so the temporal ordering of biomarkers can be stated with credible intervals rather than point estimates.","When the true curve is asymmetric, the S-shaped spline estimates milestones with lower error and better coverage than the logistic model, which shows severe inflection-point bias under misspecification.","The unconstrained monotone spline fails to plateau at early and late ages and extrapolates badly outside the observed age range, so shape constraints are needed for reliable estimation of biomarker progressions with short follow-up.","The BIOCARD estimates place CSF biomarkers first, MRI second, and cognitive biomarkers last in both milestones, consistent with the Jack model's hypothesized progression order.","Because the method only requires S-shaped trajectories, it can be applied, with modifications, to other neurodegenerative diseases whose motor, cognitive, or psychiatric signs follow similar curves."],"supporting_citations":[{"why":"Supplies the scientific hypothesis of temporal ordering and non-linear S-shaped biomarker evolution that the model's shape constraints encode.","marker":"Jack and others, 2010"},{"why":"Updates the hypothetical model and introduces the idea of a pathophysiological pathway, cited for the same ordering and for heterogeneous disease onsets.","marker":"Jack and others, 2013"},{"why":"Monotone spline regression via integrated B-splines; the starting point for the shape-constrained construction.","marker":"Ramsay, 1988"},{"why":"Shows that linear inequalities on spline coefficients enforce a unimodal regression curve, which is the basis for the unique-inflection-point constraint.","marker":"Köllmann and others, 2014"},{"why":"Penalized B-spline smoothing; the roughness penalty is used in the prior on the spline coefficients.","marker":"Eilers and Marx, 1996"},{"why":"Parametric logistic model for biomarker progression; the main comparison baseline in the simulation study.","marker":"Jedynak and others, 2012"},{"why":"Source of the BIOCARD longitudinal cohort and variable definitions used in the real-data application.","marker":"Albert and others, 2014"},{"why":"Numerical method for multivariate normal probabilities used in sampling the inflection indices and updating hyperparameters.","marker":"Genz, 1992"}],"fun_headline_variants":["Bayesian S-curves place CSF before MRI before cognition","Single-inflection splines sequence Alzheimer's biomarkers","Shape-constrained Bayesian fit ranks CSF, MRI, cognitive tests","S-shaped model timestamps AD biomarker progression","One turning point orders eleven Alzheimer biomarkers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All biomarkers within a category are forced to share the same inflection index, so within-category differences in when the fastest change occurs are assumptions built into the model rather than findings learned from the data, and no sensitivity analysis for this choice is provided.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian S-curves place CSF before MRI before cognition","Single-inflection splines sequence Alzheimer's biomarkers","Shape-constrained Bayesian fit ranks CSF, MRI, cognitive tests","S-shaped model timestamps AD biomarker progression","One turning point orders eleven Alzheimer biomarkers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2798,"prompt_tokens":953,"completion_tokens":1845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1772}},"tokens_in":569,"tokens_out":1845,"duration_ms":15990,"temperature":1.0,"reasoning_tokens":1772,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:58:50.294312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit the BIOCARD analysis with each biomarker allowed its own inflection index instead of one shared per category; if the posterior credible intervals for inflection ages within any category separate by several years, the shared-index constraint that shapes the paper's within-category ordering would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the scientific hypothesis of temporal ordering and non-linear S-shaped biomarker evolution that the model's shape constraints encode."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Updates the hypothetical model and introduces the idea of a pathophysiological pathway, cited for the same ordering and for heterogeneous disease onsets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Monotone spline regression via integrated B-splines; the starting point for the shape-constrained construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Penalized B-spline smoothing; the roughness penalty is used in the prior on the spline coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Parametric logistic model for biomarker progression; the main comparison baseline in the simulation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the BIOCARD longitudinal cohort and variable definitions used in the real-data application."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Numerical method for multivariate normal probabilities used in sampling the inflection indices and updating hyperparameters."}],"review_version":1}