{"id":"d591dc45-aed8-45c2-9d01-5da860504370","arxiv_id":"2501.09805","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A per-node decomposition of multiplex modularity is applied to working-memory binding fMRI and DTI networks and is claimed to flag MCI-to-AD converters.","lead":"Researchers define nQ, a per-brain-region version of the standard modularity score, and apply it to fMRI and DTI scans from people with mild cognitive impairment. They report regional modularity changes and claim the measure separates patients who later convert to Alzheimer's disease from those who do not.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract and Conclusion claim nQ differentiated MCI from MCI converters, but Results report no significant differences for that contrast; the central claim is unsupported by the paper's own data.","rationale":"We focused on the unsupported MCI vs. converters claim because it is a direct, checkable contradiction of the central assertion, not a methodological uncertainty. The partition-instability concern raised by the reader is real, but it is about robustness; the MCI vs. converters claim is about whether the key result exists at all. The paper's own text and S2 Table show no significant findings for that contrast, so the abstract's phrase 'clearly differentiated' is not backed by the reported statistics. This alone justifies the REJECT verdict. We keep the verdict UNCHANGED and note partial agreement with the reader: both of us flag this issue, but the reader's identified weakest assumption is the Louvain stability, which is a different (though related) limitation.","tokens_in":26229,"tokens_out":5704,"duration_ms":51187,"concrete_test":"Request or re-run the authors' pipeline for the MCI vs. MCI converters contrast on the fMRI multiplex and DTI networks, using the same permutation test with 10,000 permutations and FDR at alpha=0.2, and report the full per-ROI p-value table. If no ROI survives FDR (or if the minimum p exceeds the threshold), the abstract's claim is empirically falsified by the paper's own data; if ROIs do survive, the manuscript must be corrected to include them.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that nQ clearly differentiated MCI from MCI converters (Abstract; Conclusion) is contradicted by the Results. In the fMRI section, the authors state: 'in our comparisons of eMCI vs. MCI and MCI vs. MCI converters for fMRI shape and binding, no p-values survived FDR correction.' For DTI, S2 Table lists significant ROIs for early MCI vs. MCI (right postcentral, right precentral) but leaves the 'MCI vs. MCI converters' row empty, and no such ROIs are described in the text. Thus the paper's own statistical analysis provides no support for the headline differentiation claim. This is not a matter of interpretation: the reported p-values/AUC for the MCI vs. converter contrast are absent, and the only significant contrast reported is controls vs. MCI converters. Because the abstract explicitly touts sensitivity to the MCI-to-AD turning point, this missing result undermines the stated central claim regardless of the method's other merits.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces nodal modularity (nQ), defined as the node-level summand of the standard multislice modularity quality function, and applies it to task-fMRI and DTI multiplex networks from a small cohort performing a visual short-term memory binding task (VSTMBT). The authors benchmark nQ against other nodal measures on their own data, the NKI-Rockland cohort, and Zachary's Karate Club, and then compare regional nQ between controls, early MCI, MCI, and MCI converters. They report significant regional nQ differences in the binding task between controls and MCI converters, and claim in the Abstract and Conclusion that nQ 'clearly differentiated MCI from MCI converters', suggesting sensitivity to the MCI-to-AD turning point.","tokens_in":26391,"tokens_out":4232,"duration_ms":43877,"significance":"The mathematical definition of nQ is straightforward, the code is publicly provided, and the benchmarking against other nodal measures on multiple datasets is a useful sanity check that partially demonstrates the measure's distinct behavior. If the clinical claim were supported, nQ could offer a regional biomarker for the MCI-to-AD transition and a general tool for local community analysis. However, the central clinical claim is not supported by the reported statistics, which substantially limits the significance of the application while leaving the methodological contribution as the main value.","major_comments":[{"comment":"The Abstract and Conclusion claim that nQ 'clearly differentiated MCI from MCI converters', but the Results state that 'in our comparisons of eMCI vs. MCI and MCI vs. MCI converters for fMRI shape and binding, no p-values survived FDR correction', and S2 Table lists no ROIs for the 'MCI vs. MCI converters' contrast in the DTI networks. The only significant contrast reported is controls vs. MCI converters. Thus the paper's own data provide no evidence for the headline claim of sensitivity to the MCI-to-AD turning point; the claim is contradicted by the reported results.","section":"Results, fMRI subsection; S2 Table; Abstract; Conclusion"},{"comment":"The ROC AUC values in Tables 2 and 3 are computed on the same ROIs that were selected based on permutation-test p-values from the same data. This selection-after-testing procedure makes the AUCs optimistically biased and does not provide an independent measure of discriminative performance. Additionally, with FDR at alpha = 0.2, up to 5 of 25 fMRI ROIs and 1 of 4 DTI ROIs could be false positives, and the paper does not account for the number of group contrasts when interpreting these results.","section":"Tables 2 and 3"},{"comment":"The paper computes nQ using a single community assignment g obtained as the highest-Q partition from 100 Louvain runs, and reports only that global Q is tightly distributed across runs. It does not assess the stability of node-level community assignments or of per-node nQ across near-optimal partitions. If node assignments vary substantially among equally good partitions, group differences in nQ could reflect partition-choice artifacts rather than biological differences; this premise underlies Eq (1) and is load-bearing for the clinical comparisons.","section":"Modularity maximization"}],"minor_comments":[{"comment":"The phrase 'compliment studies' (Abstract) and 'complimenting our fMRI results' (Discussion) should be 'complement studies' and 'complementing our fMRI results', respectively.","section":"Abstract and Discussion"},{"comment":"There is a typo: 'neurospsychological examinations' should be 'neuropsychological examinations'.","section":"Introduction"},{"comment":"The notation in Eq (1) is not fully defined: the summation indices 'jsr' should be expanded, and the meaning of the product term 'δ(gis, gjr)' in the multiplex setting should be clarified for readers unfamiliar with Mucha's formulation.","section":"Eq (1) and surrounding text"},{"comment":"The decision to treat negative correlations as positive for modularity maximization is mentioned in a single sentence; given that this assumption can materially alter the community structure, it should be justified more thoroughly and its potential impact on the results discussed.","section":"Modularity maximization"},{"comment":"In the S3 Table header, 'E(2,28)' should be 'F(2,28)' for the ANOVA statistics.","section":"S3 Table"}],"recommendation":"reject","confidential_remarks":"The overstatement in the Abstract and Conclusion relative to the reported statistics is a serious issue even under a lenient reading. The central claim of MCI vs. MCI-converter differentiation is directly contradicted by the paper's own Results and S2 Table. Barring additional analyses that demonstrate a significant direct contrast, this claim should be withdrawn; the methodological contribution alone, while possibly interesting, would not carry the paper at its current level of statistical support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the headline claim is unsupported by the paper's own statistics. The abstract and conclusion say nQ 'clearly differentiated MCI from MCI converters,' but the Results state that for fMRI, no p-values survived FDR correction for eMCI vs. MCI or MCI vs. MCI converters, and S2 Table shows no significant ROIs for the DTI MCI vs. converter contrast. The only significant contrast reported is controls vs. MCI converters. That is a load-bearing mismatch between claims and evidence.\n\nWhat is genuinely new: the per-node decomposition of multislice modularity is not new—Eq. (1) is a rearrangement of Mucha et al. and is attributed to Arenas et al.—but applying it to a multiplex fMRI/DTI aMCI cohort during a memory binding task is a legitimate exploratory contribution. The code is public, the math is correct, and the benchmarking on Karate Club and NKI is a reasonable sanity check. The authors also honestly acknowledge the small sample size and the loose FDR threshold, and their discussion ties the regional findings to tau and amyloid literature without overclaiming the mechanism.\n\nSoft spots, in proportion: First, the missing MCI-vs-converter test. That is not a minor omission; it directly contradicts the abstract. Second, the group definitions look partially overlapping: Table 1 shows 16 MCI patients and the text says 6 of them converted to AD, so the MCI group appears to include the converters. Comparing MCI vs. MCI converters then uses overlapping samples, which is a serious statistical flaw. Third, the node selection procedure—FDR at alpha=0.2 followed by AUC on the same nodes—inflates apparent discrimination. Fourth, the stability of per-node nQ across Louvain runs is not checked; the authors report global Q is tightly distributed, but node assignments could vary across near-optimal partitions. That matters because nQ is computed against a single partition. These are fixable with reanalysis or a far more cautious framing.\n\nWho this is for: someone working on multiplex brain network metrics and AD biomarkers might find the application interesting as a pilot. The paper deserves a serious referee because the metric, though not new, is sound and the dataset is clinically relevant. But as written, the central claim is contradicted by the data, and the overlapping-group comparison needs to be resolved before publication.","headline":"The abstract and conclusion claim nQ clearly differentiated MCI from MCI converters, but the paper's own results show no significant contrast for that comparison; the only significant findings are controls vs. converters.","tokens_in":26974,"tokens_out":2344,"would_cite":false,"duration_ms":25610,"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 node-level modularity metric isolates the brain regions that mark MCI-to-Alzheimer's conversion.","keywords":["nodal modularity","multiplex networks","Alzheimer's disease","mild cognitive impairment","visual short-term memory binding","functional MRI","diffusion tensor imaging","community detection"],"falsifier":"Recompute nQ for every subject across many near-optimal community partitions, for example all partitions within a small modularity gap of the maximum, or across consensus partitions; if the controls-versus-converter regional differences disappear or flip sign, then the reported separation depends on the chosen partition rather than on the brain's regional community structure.","tokens_in":25997,"feed_emoji":"🧠","tokens_out":8269,"duration_ms":83810,"temperature":0.7,"pith_summary":"This paper introduces nQ, a way to measure how much each individual node contributes to a network's modularity, rather than reporting a single global number. The authors argue that this matters for Alzheimer's disease because the disease disrupts brain community structure at regional scale, and global modularity hides which regions drive the change. Applying nQ to multiplex fMRI and DTI networks from a visual short-term memory binding task, they report that the measure distinguishes MCI patients who will convert to Alzheimer's from stable MCI, with changes concentrated in visual, limbic, and paralimbic regions in the functional data and in right parietal and frontal white-matter regions in the structural data. A sympathetic reader would care because a region-specific, task-sensitive network biomarker could help identify the prodromal turning point of AD and could be transferred to other disciplines where modularity is already used as a global metric.","feed_headline":"Node-level modularity flags MCI patients who convert to Alzheimer's","feed_subtitle":"A per-region version of modularity catches brain-network shifts in a memory-binding task that global scores miss.","key_machinery":"The central object is nodal modularity, nQ, defined as the per-node summand of multislice modularity. For node $i$, nQ sums the observed-minus-expected weight of edges from $i$ to nodes in the same community, across layers and including inter-layer couplings, normalized by $2\\mu$; because these summands add to the global modularity $Q$, nQ localizes community structure to individual brain regions while preserving the standard multislice null model. The paper computes nQ from a community assignment obtained by an iterated modularity-maximization algorithm, then tests per-region group differences with permutation tests and ROC analysis.","core_discovery":"The paper claims that extending global modularity to individual nodes yields nQ, a per-node measure that localizes community structure, and that in a visual short-term memory binding task this measure detects regional brain-network reorganization in people with MCI who later convert to Alzheimer's disease. In the multiplex fMRI networks, 25 ROIs, mostly in visual, limbic, and paralimbic systems, showed abnormal nQ in converters versus controls during the binding task, while the shape-only task showed no significant effects; single-layer models found 20 ROIs. In DTI networks, nQ changes appeared in right parietal and frontal regions. The authors interpret these patterns as consistent with known amyloid-β and tau deposition and with white-matter integrity findings, and they conclude from these contrasts that nQ can differentiate MCI from MCI converters, the key prodromal turning point of AD.","pith_inferences":["Beyond the paper, because nQ is defined from any community assignment and not only from modularity maximization, it could serve as a general node-influence diagnostic for the output of any community-detection method, turning a methodological caveat into a feature.","Beyond the paper, a natural next experiment is to test nQ on EEG or MEG data, where global modularity has already been linked to AD; if nQ localizes the same effects, it could provide a cheaper and more available biomarker than fMRI.","Beyond the paper, the paper compares nQ with degree, clustering coefficient, and PageRank, but not with within-module degree or participation coefficient, the standard nodal community measures; a head-to-head comparison would clarify whether nQ adds information beyond these established metrics.","Beyond the paper, the sharp controls-versus-MCI versus controls-versus-converter contrast suggests that nQ may be more sensitive to imminent conversion than to current diagnostic category; validating this would require a prospective study with larger groups and longer follow-up."],"forward_implications":["nQ gives brain-network studies a per-region readout that sums to the global modularity they already report, so existing modularity analyses can be re-examined at node level without changing the null model.","In the paper's data, the binding task but not the shape task produced significant nQ differences, reinforcing that short-term memory binding is the AD-sensitive cognitive process and suggesting nQ reflects task-specific reorganization.","Multiplex modelling of the encoding/maintenance and probe phases found more ROIs with abnormal nQ than analysing layers separately, implying that temporal-layer coupling carries disease signal that single-layer analyses miss.","The regions flagged in fMRI overlap with sites of amyloid-β and tau deposition in the AD literature, so nQ may track molecular pathology at a regional scale.","DTI changes in right parietal and frontal ROIs, including the cuneus, link nQ to white-matter integrity and cross-hemispheric communication, suggesting a structural counterpart to the functional reorganization."],"supporting_citations":[{"why":"Defines the visual short-term memory binding task and establishes its sensitivity to Alzheimer's disease; the task is the paper's cognitive paradigm.","marker":"[7]"},{"why":"Supplies the multislice modularity quality function from which nQ is derived.","marker":"[25]"},{"why":"Provides the size-reduction derivation used to split modularity into per-node contributions.","marker":"[41]"},{"why":"Provides the iterated community-detection algorithm used to obtain the community assignment g on which nQ depends.","marker":"[42]"},{"why":"Establishes that global modularity changes along the AD spectrum, motivating a node-level version.","marker":"[26]"},{"why":"Provides the fMRI task-phase design and ROI analyses of shape-color binding underlying the network layers.","marker":"[27]"},{"why":"Describes the longitudinal MCI cohort and diagnostic grouping, including converters, that the study analyzes.","marker":"[28]"},{"why":"Reports amyloid-β deposition in regions that align with the nQ changes observed in poor memory binders.","marker":"[66]"}],"fun_headline_variants":["Nodal modularity flags MCI patients who later get Alzheimer's","Per-node network metric spots Alzheimer's-bound MCI","Local modularity in brain nets predicts MCI conversion","New nodal modularity metric: sees MCI turning into AD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis uses one community assignment, the best of 100 modularity-maximization runs, as a fixed backdrop for every node's nQ, and if nodes switch communities between nearly equivalent partitions, the group differences could be a byproduct of that choice.","fun_headline_variants_meta":{"raw":{"variants":["Nodal modularity flags MCI patients who later get Alzheimer's","Per-node network metric spots Alzheimer's-bound MCI","Local modularity in brain nets predicts MCI conversion","New nodal modularity metric: sees MCI turning into AD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1587,"prompt_tokens":1063,"completion_tokens":524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":679,"tokens_out":524,"duration_ms":5414,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:39:28.959893+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute nQ for every subject across many near-optimal community partitions, for example all partitions within a small modularity gap of the maximum, or across consensus partitions; if the controls-versus-converter regional differences disappear or flip sign, then the reported separation depends on the chosen partition rather than on the brain's regional community structure.","supporting_citations":[{"cited_title":"Visual short-term memory binding deficits in familial Alzheimer’s disease","cited_arxiv_id":null,"evidence_quote":"Defines the visual short-term memory binding task and establishes its sensitivity to Alzheimer's disease; the task is the paper's cognitive paradigm."},{"cited_title":"Size reduction of complex networks preserving modularity","cited_arxiv_id":null,"evidence_quote":"Provides the size-reduction derivation used to split modularity into per-node contributions."},{"cited_title":"Disrupted Network Topology in Patients with Stable and Progressive Mild Cognitive Impairment and Alzheimer’s Disease","cited_arxiv_id":null,"evidence_quote":"Establishes that global modularity changes along the AD spectrum, motivating a node-level version."},{"cited_title":"Neural correlates of shape-color binding in visual working memory","cited_arxiv_id":null,"evidence_quote":"Provides the fMRI task-phase design and ROI analyses of shape-color binding underlying the network layers."},{"cited_title":"Memory markers in the continuum of the Alzheimer’s clinical syndrome","cited_arxiv_id":null,"evidence_quote":"Describes the longitudinal MCI cohort and diagnostic grouping, including converters, that the study analyzes."}],"review_version":1}