{"id":"d5fbad10-36a6-4a9a-bef5-ef21ff6b36db","arxiv_id":"1908.05627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A symmetric bilinear logistic regression with elastic-net penalty learns time-invariant clique signal subgraphs with age-varying weights from longitudinal brain networks, and on ADNI data selects a frontal-temporal pole subgraph with constant age effect.","lead":"This paper proposes a logistic regression model that finds small clique-shaped groups of brain connections predicting 'supernormal' cognitive aging from repeated MRI scans. On 85 ADNI participants it reports lower cross-validated prediction error than a standard penalized regression and selects a frontal-temporal pole subgraph.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's claim of better predictive performance rests on two unquantified, possibly leaky CV point estimates; this is the load-bearing weakness.","rationale":"I read the paper as a methods contribution whose abstract makes a comparative empirical claim: SBLR gives 'better predictive performance than competitors' on ADNI data. The reader's weakest assumption concerns temporal stability of signal subgraphs; that is a real limitation but it is explicit in Section 2 and is a modeling choice the paper can be judged under. The more directly load-bearing gap is the evidence for the superiority claim: Section 5 reports single mean CV deviance values with no uncertainty, and the preprocessing in Section 3 appears to standardize the full dataset before the CV split described in Section 3.5. If so, the CV protocol leaks held-out information, and even without leakage the absence of variance estimates means a small n=85 difference cannot be evaluated. This does not invalidate the method: the simulation study, the worked update equations, and the public code are real support. But it does mean the headline empirical claim is not yet established. The reader's CONDITIONAL verdict is therefore appropriate; I would keep it unchanged rather than moving to reject, because the concern is addressable by a corrected repeated-CV analysis. I agree partially with the reader: they flagged normalization and missing standard errors in the rationale, but their stated weakest assumption was the temporal-stability condition, which I see as secondary to the CV evidence problem.","tokens_in":18900,"tokens_out":12900,"duration_ms":143478,"concrete_test":"Re-run the FA and MD analyses with a corrected protocol: perform edgewise standardization of W, g, and g^2 inside each training fold only, apply the same transforms to the left-out fold, and repeat 5-fold CV across 50 random fold partitions. Report the mean and SD of the SBLR-minus-LR CV deviance difference. If the mean advantage is smaller than its SD, or if fold-aware normalization shrinks SBLR's advantage by more than 0.05, the paper should remove or soften the 'better predictive performance' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing condition for the headline claim is that the Section 5 comparison actually demonstrates better predictive performance. The manuscript reports only point estimates of mean CV deviance: SBLR 1.29 vs LR 1.38 for FA, and 1.31 vs 1.39 for MD. No fold-level or repeated-split variability is given, so with n=85 and balanced classes a 0.09-0.10 difference is within plausible sampling noise. The problem is compounded by the preprocessing order: Section 3 states that W entries and age terms are standardized over 'the observations' before model fitting, and Section 3.5 then performs 5-fold CV for tuning. If that normalization was computed on the pooled data before the folds were formed, the edgewise means and variances of held-out subjects enter the training predictors, making the CV deviance optimistically biased. The paper does not state that normalization is fold-aware. Since the abstract's 'better predictive performance than competitors' rests on these two comparisons, this is the weakest link in the central claim. The temporal-stability assumption in Section 2 is explicit and is a modeling choice; the CV evidence is presented as empirical support and is currently unquantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes symmetric bilinear logistic regression (SBLR) for binary outcomes from longitudinal brain networks. The model assumes that a small number of outcome-relevant signal subgraphs are captured by component vectors β_h, and that the predictive effect of each subgraph evolves with age through a shared quadratic function λ_h(g) = γ_h g^2 + ρ_h g + α_h. An elastic-net penalty on the lower-triangular entries of λ_h(g) β_h β_h^T encourages sparse clique subgraphs. Estimation is performed by a coordinate descent algorithm with quadratic approximations, and model selection uses 5-fold cross-validation with a one-standard-error rule. Simulations compare SBLR with unstructured logistic regression and a naively symmetrized tensor regression. An application to ADNI data with FA and MD connectivity matrices identifies a frontal/temporal-pole subgraph and reports lower mean CV deviance for SBLR than for logistic regression.","tokens_in":19091,"tokens_out":4041,"duration_ms":37927,"significance":"If the claims hold, the paper offers a useful structured alternative to edge-wise logistic regression for longitudinal network predictors, with interpretable clique subgraphs and population-level age effects. The coordinate-wise update derivations are explicit, the simulation protocol is sensible, and the authors provide publicly available code, which are strengths. The method is an incremental but nontrivial extension of Wang et al. (2019) to binary longitudinal outcomes with age-varying effects. However, the empirical support for the headline claim of \"better predictive performance than competitors\" is currently thin: the reported CV deviance differences in the application are unquantified, the preprocessing may leak information into cross-validation, and the convergence guarantee is asserted rather than proved. These issues are fixable but need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The abstract and Section 5 claim \"better predictive performance than competitors\" based on point estimates of mean CV deviance: SBLR 1.29 versus LR 1.38 for FA, and SBLR 1.31 versus LR 1.39 for MD. With n = 85 and balanced classes, these differences are within plausible sampling noise, and no fold-level variability, repeated-split estimates, or paired comparison is provided. Report standard errors or confidence intervals for the CV deviance (or a paired test across folds) and temper the predictive-performance claim accordingly.","section":"§5.1 and §5.2"},{"comment":"The preprocessing described at the start of Section 3 standardizes the raw network entries and age terms over \"the observations\" before model fitting, and Section 3.5 then performs 5-fold CV for tuning. If this standardization is computed on the pooled data before the folds are formed, held-out subjects' edge means and variances leak into the training predictors, biasing the CV deviance optimistically. Clarify whether normalization is fold-aware; if it is not, repeat the evaluation with within-training-fold standardization and report whether the conclusions change.","section":"§3 and §3.5"},{"comment":"The convergence guarantee is asserted rather than proved: \"Since the loss function (5) is lower bounded by 0 and each update always decreases the function value, the coordinate descent algorithm derived above is guaranteed to converge.\" Each update in equations (17)–(26) minimizes a second-order Taylor approximation of the loss, not the true loss, so monotone decrease of the true objective is not automatic. Provide a descent lemma or modify the algorithm (for example, with an exact coordinate minimization or a line search) to make the convergence claim rigorous.","section":"§3.4"},{"comment":"The simulation evidence for signal-subgraph recovery is weaker than the narrative suggests: at n = 100 the mean TPR for SBLR is 0.219 (K = 5) with SD 0.271, and the displayed run recovers only part of one of the two true subgraphs. The conclusion that SBLR \"learns\" signal subgraphs should be framed as partial recovery with strong false-positive control, or additional simulation configurations with stronger signal-to-noise ratio should be considered to substantiate the recovery claim.","section":"§4.2, Table 1"}],"minor_comments":[{"comment":"The word \"summerized\" in Algorithm 1's description should be \"summarized\".","section":"§3.4"},{"comment":"The middle panel says \"red edges ... (not displayed)\"; this parenthetical contradicts the legend and should be removed or the red edges should be shown.","section":"§4.2, Figure 8"},{"comment":"The definition of δ_min as \"a sufficiently small value that produces dense results\" is vague; give an operational definition or a default formula for choosing the grid endpoints.","section":"§3.5"},{"comment":"The identifiability of the elastic-net penalty is motivated by a reference to Wang et al. (2019); a brief self-contained statement of the identifiability argument for the logistic and longitudinal setting would help readers.","section":"§2"},{"comment":"In the application, SBLR is compared only with LR, not with NSTR or other structured competitors; the abstract's plural \"competitors\" overstates the scope of the empirical comparison.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The methodology is potentially publishable, but the application claim needs statistical support: the CV deviance comparison should be accompanied by variability estimates and a fold-aware preprocessing description. As an editor, I would ask for a revision addressing these issues before considering acceptance. There is also a scope-fit question: for a statistics journal with an applications focus, the real-data evaluation should meet the same evidentiary standard as the methodological sections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a logistic regression extension of the authors' earlier symmetric bilinear model, now handling longitudinal network predictors and age-varying coefficients. The coordinate descent derivations are explicit, the per-iteration complexity analysis is honest, and the sparsity/interpretability benefit over unstructured elastic-net regression comes through in the simulations. Code is on GitHub, which helps. The soft spots are real but mostly addressable. The biggest one is the application section. The abstract says SBLR beats competitors on ADNI, but the evidence is only two point estimates of mean CV deviance (1.29 vs 1.38 for FA; 1.31 vs 1.39 for MD) with n=85. No standard errors, no repeated splits. A difference of 0.08–0.10 is well within plausible sampling noise at that sample size. Worse, Section 3 describes normalizing the edge weights and age terms on the observations before model fitting, and Section 3.5 then does 5-fold CV. If that normalization was computed on pooled data, the fold structure is breached and the CV deviance is optimistically biased. The paper never states that normalization is fold-aware. That is the load-bearing weakness in the headline claim. Two smaller things. First, the convergence guarantee is asserted rather than proven: the algorithm uses a quadratic approximation in coordinate descent, and while the loss is lower bounded, that alone does not guarantee each update decreases it. A line search or a proper majorization argument would settle it. Second, Table 1 contains an unexplained value: under n=1000, SBLR has mean CV deviance 0.1767 while LR and NSTR are around 1.18. That looks like a typo, and it makes the whole table harder to trust. The temporal stability assumption on the subgraphs is explicit and reasonable as a first model, though it should be flagged more prominently as a limitation in the discussion. The methodological core is sound and the simulation protocol is sensible. The paper is an incremental but useful tool for connectomics researchers who want sparse, interpretable subgraph regression for binary outcomes. It deserves a serious referee. My recommendation: send it to peer review, but the application section needs fold-aware preprocessing, uncertainty quantification on the CV deviance, and a corrected Table 1 before publication.","headline":"A solid methodological extension with a credible derivational core, but the headline ADNI predictive claim is under-quantified and may be leaky; referee it, but expect revisions.","tokens_in":652,"tokens_out":789,"would_cite":false,"duration_ms":28041,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A symmetric bilinear logistic regression learns small clique subgraphs from longitudinal brain networks and estimates how each subgraph's predictive weight changes with age; on ADNI connectomes it identifies a frontal- and temporal-pole…","keywords":["signal subgraph learning","longitudinal brain networks","symmetric bilinear logistic regression","clique subgraphs","age effect","structural connectomics","elastic-net penalty","supernormal aging"],"falsifier":"Refit or evaluate the model on age-banded subsets of the ADNI data, for example subjects seen near ages 60-70 versus 80-95, allowing the subgraph membership $\\beta_h$ to be re-estimated within each band. If the same frontal- and temporal-pole clique does not predict supernormal status in every band, or if a different clique is selected, the fixed-subgraph assumption is violated and the reported age effect is an artifact of averaging.","tokens_in":18666,"feed_emoji":"🧠","tokens_out":10194,"duration_ms":91952,"temperature":0.7,"pith_summary":"This paper proposes a way to predict a binary trait from a set of longitudinal brain networks, treating each network as a matrix predictor rather than a bag of edges. The model, symmetric bilinear logistic regression (SBLR), represents each outcome-relevant subgraph as a rank-one symmetric matrix $\\beta_h \\beta_h^\\top$, so the selected connections form a clique—a small set of mutually connected brain regions that can be read as a neurological circuit. A scalar function $\\lambda_h(g) = \\gamma_h g^2 + \\rho_h g + \\alpha_h$ multiplies each clique and lets its predictive weight change with age while the subgraph itself stays fixed. On ADNI structural connectomes, SBLR identifies a frontal-pole-centered clique whose connection strength is positively associated with \"supernormal\" cognitive aging, and it reports lower cross-validated deviance than an unstructured logistic-regression baseline. The payoff, if correct, is that a high-dimensional edge-selection problem collapses to a few interpretable circuits with estimable age trajectories.","feed_headline":"Frontal-pole clique spots 'supernormal' aging from brain scans","feed_subtitle":"One regression distills years of brain scans into a frontal-pole circuit with an estimated age effect.","key_machinery":"The central object is the rank-one symmetric coefficient matrix $\\beta_h \\beta_h^\\top$ inside a logistic link. Its nonzero off-diagonal entries mark the edges of a clique subgraph, so each component selects a set of mutually connected nodes that can be interpreted as a circuit; the age effect is carried by a scalar polynomial $\\lambda_h(g) = \\gamma_h g^2 + \\rho_h g + \\alpha_h$ multiplying the whole clique. With $\\lambda_h$ quadratic, the longitudinal predictor collapses into three per-subject matrix averages: $\\frac{1}{T_i} \\sum_s W_i^{(s)}$, $\\frac{1}{T_i} \\sum_s g_{is} W_i^{(s)}$, and $\\frac{1}{T_i} \\sum_s g_{is}^2 W_i^{(s)}$, which is what lets the method handle subjects with different numbers of visits. Estimation uses coordinate descent with an entrywise elastic-net penalty whose per-entry weight grows with the current $\\ell^1$ mass of the rest of the component, pushing unnecessary components to zero and keeping the selected subgraphs small; the loss is not jointly concave in $\\beta_h$, but each coordinate update is convex.","core_discovery":"On the paper's own terms, the central discovery is that a logistic model with components $\\mathrm{logit}(p_i) = \\alpha_0 + \\sum_{h=1}^K \\frac{1}{T_i} \\sum_{s=1}^{T_i} \\lambda_h(g_{is}) \\beta_h^\\top W_i^{(s)} \\beta_h$ can simultaneously select sparse clique subgraphs and estimate their age effects from longitudinal networks with irregular visit schedules. Because each component's coefficient matrix is $\\lambda_h(g) \\beta_h \\beta_h^\\top$, nonzero off-diagonal entries sit exactly on a clique, and the shared quadratic $\\lambda_h(g)$ captures how that circuit's contribution to the outcome drifts with age. Simulations show the method recovers true signal edges with the lowest false-positive rate among SBLR, unstructured elastic-net logistic regression, and a tensor-regression baseline, while matching or beating them in out-of-sample deviance. In the ADNI application, SBLR selects one clique around the frontal pole, with superior frontal and temporal pole nodes, whose connection strengths are positively and constantly associated with supernormal status across ages, with mean cross-validated deviance 1.29 (FA) and 1.31 (MD) versus 1.38 and 1.39 for the unstructured logistic regression.","pith_inferences":["A stronger test of the paper's stability assumption would let $\\beta_h$ vary by age window, for example 60-74 versus 75-95; the current model only varies $\\lambda_h$, so the inference that the same circuit operates throughout aging is not directly tested.","The identified frontal-pole-centered circuit generates a concrete prediction for independent cohorts: white-matter integrity among frontal pole, superior frontal, and temporal pole regions should track memory performance in older adults classified as superagers.","The per-subject averaging trick could be reused for functional connectomes, such as fMRI, after aligning visits to age bins, making the same logistic-bilinear construction applicable beyond diffusion MRI."],"forward_implications":["A handful of clique subgraphs, each tied to named anatomy, can replace long edge lists in longitudinal brain-outcome studies.","Subjects with different numbers of visits can be pooled without imputation, because the model uses per-subject averages of the network, the age-weighted network, and the age-squared-weighted network.","The estimated $\\hat{\\lambda}_h(g)$ gives a population-level trajectory of how a circuit's predictive weight changes with age; a constant trajectory means no age window is especially decisive.","The same clique emerges from fractional anisotropy and mean diffusivity connectomes, suggesting the finding is not an artifact of one diffusion measure.","SBLR is conservative relative to unstructured elastic-net regression, favoring fewer false-positive edges at comparable predictive deviance."],"supporting_citations":[{"why":"Introduces symmetric bilinear regression for clique signal subgraph estimation, the model that SBLR adapts to binary longitudinal outcomes.","marker":"Wang et al. (2019)"},{"why":"Supplies the coordinate-descent updates for regularized generalized linear models and the one-standard-error cross-validation rule used for tuning.","marker":"Friedman et al. (2010)"},{"why":"Defines the tensor-regression baseline and the block relaxation algorithm that SBLR shows is inapplicable because of non-concavity.","marker":"Zhou et al. (2013)"},{"why":"Provides the connectome extraction pipeline that turns diffusion MRI data into the structural brain networks used in the ADNI analysis.","marker":"Zhang et al. (2018)"},{"why":"Defines supernormals and supplies the ADNI subject selection; the paper's outcome is prediction of this supernormal status.","marker":"Lin et al. (2017)"},{"why":"Defines the 68-region cortical atlas whose node labels identify the frontal and temporal pole regions in the discovered subgraph.","marker":"Desikan et al. (2006)"}],"fun_headline_variants":["One regression finds frontal-pole circuit behind 'supernormal' aging","Sparse clique in frontal lobe marks supernormal brain aging","Bilinear regression distills aging brain into one useful circuit","Frontal-pole clique ties brain scans to 'supernormal' aging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the brain regions and their connections in each signal subgraph stay the same for everyone across all ages; the model only lets the statistical weight of each fixed circuit change with age.","fun_headline_variants_meta":{"raw":{"variants":["One regression finds frontal-pole circuit behind 'supernormal' aging","Sparse clique in frontal lobe marks supernormal brain aging","Bilinear regression distills aging brain into one useful circuit","Frontal-pole clique ties brain scans to 'supernormal' aging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3204,"prompt_tokens":1006,"completion_tokens":2198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2126}},"tokens_in":622,"tokens_out":2198,"duration_ms":15400,"temperature":1.0,"reasoning_tokens":2126,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:07:59.995417+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit or evaluate the model on age-banded subsets of the ADNI data, for example subjects seen near ages 60-70 versus 80-95, allowing the subgraph membership $\\beta_h$ to be re-estimated within each band. If the same frontal- and temporal-pole clique does not predict supernormal status in every band, or if a different clique is selected, the fixed-subgraph assumption is violated and the reported age effect is an artifact of averaging.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces symmetric bilinear regression for clique signal subgraph estimation, the model that SBLR adapts to binary longitudinal outcomes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the connectome extraction pipeline that turns diffusion MRI data into the structural brain networks used in the ADNI analysis."},{"cited_title":"P., Porsteinsson, A., Baran, T","cited_arxiv_id":null,"evidence_quote":"Defines supernormals and supplies the ADNI subject selection; the paper's outcome is prediction of this supernormal status."},{"cited_title":"S., S \\'e gonne, F., Fischl, B., Quinn, B","cited_arxiv_id":null,"evidence_quote":"Defines the 68-region cortical atlas whose node labels identify the frontal and temporal pole regions in the discovered subgraph."}],"review_version":1}