{"id":"63ef17dc-80ee-4190-98b0-8bae9e0a395e","arxiv_id":"2607.04702","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Risk-adjusted Word2Vec-plus-GGM comorbidity networks from UK Biobank trajectories recover clinically coherent cardiometabolic modules and four progression phenotypes with divergent survival.","lead":"A new pipeline builds cardiometabolic disease networks from UK Biobank trajectories by embedding events, stripping shared-risk confounding, and estimating sparse direct links with graphical models. The resulting communities yield four patient progression types with clearly different survival.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The Δ-as-confounding step is the load-bearing soft spot: if it mainly tracks embedding geometry rather than true shared-risk confounding, the “risk-adjusted” GGM and survival phenotypes are mis-specified.","rationale":"The reader correctly isolates the weakest link: the heuristic identification of confounders via embedding-similarity change. All subsequent claims—sparse risk-adjusted network, clinically coherent communities, and survival-separated progression phenotypes—depend on that step being a faithful measure of confounding rather than an embedding artefact. The paper’s internal checks (systematic negative MDC, Mann–Whitney link to SharedRF, robustness to τ) are supportive but correlational; they do not demonstrate that conditioning on C*_i removes the intended bias or recovers true conditional independencies. A controlled simulation with known edges and known confounders is the minimal decisive test. Because the rest of the pipeline is carefully executed, code is public, and the clinical patterns are plausible, the appropriate stance remains CONDITIONAL rather than REJECT: the contribution is usable once the confounding proxy is stress-tested. No stronger internal inconsistency is present, and the reader’s other caveats (normality of embeddings, unadjusted KM curves) are secondary once this assumption is secured.","tokens_in":17686,"tokens_out":737,"duration_ms":6562,"concrete_test":"Generate synthetic longitudinal sequences from a known sparse disease graph plus a known set of shared risk factors that induce spurious co-occurrence; run the full pipeline (Word2Vec fine-tune → Δ → C*_i → GGM) and report precision/recall of recovered edges against the ground-truth graph, both with and without the C*_i covariates. If edge recovery does not improve (or worsens) when C*_i are included, the Δ-as-confounding claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on treating signed cosine-similarity change Δ_ij = A^(D+R)_ij − A^(D)_ij (Eq. 3.4, §3.3.2–3.3.4) as a quantitative proxy for confounding by shared risk factors, so that the aggregated sets C*_i can be inserted as covariates into the neighbourhood Lasso (Eq. 3.7) and thereby produce a debiased sparse network. The paper supplies only an indirect validation: disease pairs in the 95th percentile of |Δ| share more risk factors (Mann–Whitney p = 2.57e-14) and the result is robust to τ. That association is necessary but not sufficient. Because both embedding spaces are learned from the same patients, Δ can arise from (i) redistribution of co-occurrence mass onto risk-factor tokens, (ii) changes in sequence length/composition, or (iii) geometric artefacts of averaging multi-word MeSH tokens—none of which equals classical confounding. If those artefacts dominate, the “prior clinical knowledge” is misspecified, the GGM edges are not risk-adjusted, and the four survival-separated progression phenotypes inherit the same bias. The absence of a simulation with known ground-truth edges and known confounders leaves this possibility untested.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a pipeline for population-level comorbidity network inference from longitudinal UK Biobank trajectories. Individual event sequences (24 cardiometabolic diseases + 76 risk factors) are embedded via fine-tuned BioWordVec; signed changes in pairwise cosine similarity between disease-only and disease+risk embeddings (Δ, Eq. 3.4) are used to identify disease-specific confounder sets C*_i (\"prior clinical knowledge\"); these sets are inserted as covariates into neighbourhood-selection GGMs with Lasso (Eq. 3.7) to estimate a sparse risk-adjusted network. Topological analysis yields four communities aligned with cardiometabolic taxonomy; community-transition matrices are clustered into four progression phenotypes that show significantly different Kaplan–Meier all-cause survival (mortality 11.7%–32.1%). The authors report internal checks (Wilcoxon on Δ, Mann–Whitney linking |Δ| to SharedRF, τ sensitivity) and release code with simulated data.","tokens_in":18068,"tokens_out":1478,"duration_ms":11364,"significance":"If the pipeline is valid, it offers a concrete, reproducible alternative to cross-sectional co-occurrence networks that (i) uses full trajectories, (ii) attempts explicit risk-factor adjustment, and (iii) produces sparse conditional-dependence graphs rather than thresholded fully connected graphs. The downstream community-transition phenotypes with external survival separation are a useful demonstration of clinical utility. Strengths include open code, multi-source phenotype definitions, and transparent sensitivity analyses on τ. The work is of clear interest to statistical applications in multimorbidity and network medicine, provided the central confounding-identification step can be better justified.","major_comments":[{"comment":"The load-bearing claim that Δ_ij = A^(D+R)_ij − A^(D)_ij (Eq. 3.4, §3.3.2–3.3.4) quantifies confounding by shared risk factors is only indirectly supported. The Mann–Whitney result (pairs in the 95th percentile of |Δ| share more risk factors, p = 2.57e-14) and τ robustness (Supp. Table S3) show association, not that Δ isolates classical confounding rather than embedding geometry, sequence-composition change, or multi-word MeSH averaging artefacts. Because both embedding spaces are learned from the same patients, redistribution of co-occurrence mass onto risk-factor tokens can produce negative Δ without corresponding to a causal confounder. This step defines the C*_i sets that enter the neighbourhood Lasso (Eq. 3.7); if it is misspecified, the \"risk-adjusted\" edges and the four survival phenotypes inherit the bias. A simulation with known ground-truth edges and known confounders (or an ab","section":"§3.3.2–3.3.5, Eq. (3.4), Eq. (3.7)"},{"comment":"No quantitative comparison is reported between the risk-adjusted GGM network and a disease-only GGM (or a standard co-occurrence / partial-correlation baseline). Table 1 and Figure 2 characterise similarity-matrix change, but the final sparse adjacency (Figure 3) is presented only in its adjusted form. Without edge-set overlap, precision/recall against a clinical reference, or a side-by-side topological comparison, it is impossible to assess how much the prior-knowledge step actually changes the estimated network or the downstream communities and survival clusters. This comparison is required to substantiate the claim that the pipeline \"addresses\" confounding.","section":"§4, Figure 3, Table 1"},{"comment":"The multivariate Gaussian assumption for embedding vectors (treated as independent observations across dimensions) is stated and checked only via per-dimension Shapiro–Wilk tests (§3.4). Embeddings from skip-gram models are typically non-Gaussian and dimensions are not independent observations of a single multivariate law. The paper should either (i) report the actual Shapiro–Wilk outcomes and any remedial transforms, or (ii) justify neighbourhood selection under weaker assumptions (e.g., nonparanormal / copula GGMs). As written, the statistical foundation of the sparse network is under-specified.","section":"§3.4"}],"minor_comments":[{"comment":"Typo in Results opening: \"Staring from embeddings\" should be \"Starting from embeddings\".","section":"§4"},{"comment":"Supplementary Tables S1–S2 are linked only as external Excel files; a short in-text summary of the 24 diseases and main risk-factor categories would help readers who cannot access the repository immediately.","section":"§2, Supplementary Material"},{"comment":"The stop criterion sim_preserved < 0.70 after 4 epochs (§3.2) is ad hoc; a brief sensitivity check on embedding dimension d or window c would strengthen reproducibility claims.","section":"§3.2"},{"comment":"Figure 2 colour scale and caption are clear, but the adjacency matrix in Figure 3b would benefit from community-ordered rows/columns to make modular structure visible without referring to Figure 4.","section":"Figures 2–4"},{"comment":"Kaplan–Meier analysis (§3.5.2, Figure 5) reports visual separation and mortality rates but no log-rank or Cox p-value in the main text; adding the test statistic would make the survival claim fully self-contained.","section":"§4, Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid applied contribution with open code and a clear clinical narrative, but the Δ-as-confounding step is the soft underbelly and is currently validated only by association. I would not reject on that basis alone if the authors can supply a simulation or a strong ablation; without it the \"risk-adjusted\" claim remains aspirational. Fit for a methods-oriented applied statistics / biostatistics venue is good once the major points are addressed. No concerns about citation pattern or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, usable methods paper for people who build comorbidity networks from EHR. The real contribution is not any single algorithm—Word2Vec on sequences, BioWordVec fine-tuning, Lasso GGMs, and modularity are all known—but the full pipeline that turns longitudinal UK Biobank trajectories into a sparse 24-disease network, four taxonomy-aligned communities, and four patient progression phenotypes with clear Kaplan–Meier separation (mortality 11.7 %–32.1 %). Code and simulated data are public, which already puts it ahead of most of the literature it cites.\n\nWhat works: they target three genuine problems (no time, no confounding control, dense pairwise graphs) and produce something clinically readable—renal progression path, valvular module, ischemic-arrhythmic-HF triad, ESRD as the top hub. Survival is an external endpoint, so the phenotype claim is not circular. Internal checks (Wilcoxon on Δ, Mann–Whitney linking |Δ| to shared risk-factor burden, τ sensitivity) are coherent and the modules match known cardiometabolic taxonomy.\n\nThe soft spot the stress-test flags is real but proportionate. Treating signed cosine change Δ between disease-only and disease+risk embeddings as a quantitative confounding measure is a heuristic, not a formal identification result. Both spaces are learned from the same patients, so Δ can also reflect sequence composition or embedding geometry. They only show that high-|Δ| pairs share more risk factors; that is necessary, not sufficient. No simulation with known edges and known confounders is provided (they acknowledge this). GGM-on-embeddings also leans on normality/independence that is only lightly checked. None of this collapses the paper; it means the “risk-adjusted direct network” claim is better read as “confounder-informed sparse network that looks clinically sensible” until someone stress-tests the Δ step.\n\nWho it is for: applied statisticians and cardiometabolic epidemiologists who need a reproducible trajectory-to-network pipeline and a concrete risk-stratification use case. It deserves a serious referee, not a desk reject. I would bring it to reading group and would cite the pipeline and the four phenotypes if I were working in this space. Recommend peer review with a request for clearer language around the heuristic nature of Δ and, if feasible, a small simulation or stronger baseline comparison.","headline":"Solid applied pipeline that actually ships a usable risk-adjusted cardiometabolic network and survival-separated phenotypes; the Δ-as-confounding step is heuristic but not fatal.","tokens_in":18675,"tokens_out":578,"would_cite":true,"duration_ms":5526,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Risk-adjusted disease networks from health trajectories yield four multimorbidity phenotypes with sharply different survival.","keywords":["comorbidity network","Gaussian Graphical Models","Word2Vec embeddings","cardiometabolic diseases","longitudinal trajectories","confounding adjustment","multimorbidity phenotypes","UK Biobank"],"falsifier":"A controlled simulation or external cohort in which the true disease graph and true confounders are known: if the Δ-based confounder sets fail to recover the known confounders, or if the GGM edges recovered after adjustment still contain the known spurious links, the central confounding-identification claim fails.","tokens_in":18549,"feed_emoji":"📈","tokens_out":666,"duration_ms":7173,"temperature":0.7,"pith_summary":"Comorbidity networks are usually built from snapshot co-occurrence counts. That approach ignores order of disease onset, confounds associations with shared risk factors, and produces dense graphs that mix direct and indirect links. This paper builds a pipeline that starts from full patient event sequences, embeds them so that temporal and semantic proximity become geometry, then measures how much each disease-pair similarity shrinks once risk factors are jointly embedded. Those shrinkages identify disease-specific confounder sets that are fed into a Lasso Gaussian graphical model, producing a sparse network of only direct conditional dependencies. Applied to UK Biobank cardiometabolic data, the network recovers four clinically coherent communities. Patient trajectories are then rewritten as transitions among those communities; clustering the transition fingerprints yields four progression phenotypes whose all-cause survival curves separate after age 60, with mortality ranging from roughly 12 % to 32 %. The claim is that this risk-enriched, trajectory-based route both cleans the network and supplies a usable stratification tool for multimorbidity care.","feed_headline":"Four multimorbidity phenotypes emerge with different survival","feed_subtitle":"Risk-adjusted trajectory networks turn cardiometabolic histories into survival strata","key_machinery":"Prior clinical knowledge constructed as disease-specific confounder sets C*_i: the union, for each disease, of risk factors that co-associate with it in the joint embedding space and that drive large negative similarity changes Δ_ij; these sets are included as covariates in each neighbourhood-selection Lasso regression so that the estimated edges are conditional on shared risk burden.","core_discovery":"Embedding complete disease-and-risk trajectories, quantifying confounding via the signed drop in cosine similarity after risk adjustment, and feeding the resulting disease-specific confounder sets into a Lasso Gaussian graphical model recovers a sparse cardiometabolic network whose four communities support patient transition representations that cluster into four progression phenotypes with significantly different long-term survival.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Risk-adjusted trajectories yield four cardiometabolic survival phenotypes","Sparse trajectory networks reveal four multimorbidity survival strata","Four progression phenotypes emerge from confounder-adjusted disease networks","Health trajectory embeddings define four distinct multimorbidity survival groups","Community-based patient representations cluster into four survival phenotypes"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The signed change in pairwise cosine similarity between disease-only and disease-plus-risk embeddings is treated as a quantitative measure of true confounding by shared risk factors; if that geometric change mainly reflects embedding artefacts rather than confounding, the adjusted edges and the downstream phenotypes are misspecified.","fun_headline_variants_meta":{"raw":{"variants":["Risk-adjusted trajectories yield four cardiometabolic survival phenotypes","Sparse trajectory networks reveal four multimorbidity survival strata","Four progression phenotypes emerge from confounder-adjusted disease networks","Health trajectory embeddings define four distinct multimorbidity survival groups","Community-based patient representations cluster into four survival phenotypes"]},"model":"grok-4.5","effort":"low","cost_usd":0.003238,"raw_usage":{"total_tokens":1127,"prompt_tokens":785,"num_sources_used":0,"completion_tokens":79,"cost_in_usd_ticks":32380000,"prompt_tokens_details":{"text_tokens":785,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":263,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":785,"tokens_out":79,"duration_ms":3181,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T14:54:08.446387+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A controlled simulation or external cohort in which the true disease graph and true confounders are known: if the Δ-based confounder sets fail to recover the known confounders, or if the GGM edges recovered after adjustment still contain the known spurious links, the central confounding-identification claim fails.","supporting_citations":[],"review_version":1}