{"id":"4dd4ad9c-67ba-44a7-87ca-b332cb21f6f1","arxiv_id":"1908.09116","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Applying the Network Dependency Index to resting-state functional connectomes yields stable subnetworks and reveals autism-control differences in the most important subnetworks.","lead":"The authors applied a network analysis tool, the Network Dependency Index, to brain scans of people with and without autism, and found that the most important brain subnetworks show connectivity differences between the groups. The paper shows this subnetwork approach works on functional MRI data and could help identify autism-related brain changes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Group-derived NDI Tier labels are applied to individual connectomes without validation; if individual NDI topology deviates, reported ASD differences may reflect label mismatch.","rationale":"The paper's feasibility claim is supported by the RBO-based stability comparisons, and the use of a public dataset is a strength. The weakest point is the interpretive bridge from group-connectome Tier labels to individual-subject Tier measures. Section 2.2 builds group connectomes by 90% edge presence and averaging; Section 2.4 applies the resulting labels to each subject. Because NDI is a nonlinear, deletion-based measure, a group-averaged topology need not have the same hierarchy as the typical individual topology. The stability result across cohort, control, and patient group connectomes (Section 3, Table 2) speaks only to the group level. The conclusion's phrasing, that differences are located in nodes with highest importance, requires the labels to be valid at the individual level. The authors' Discussion statement that individual-level NDI subnetwork definition is future work is an explicit admission that this assumption is unvalidated. This is more load-bearing than the multiple-comparison issue: even with perfect correction, an unvalidated label transfer would make significant Tier differences hard to interpret. Multiple testing remains a secondary concern and is part of why the conditional verdict is appropriate. A concrete per-subject label-agreement analysis would settle whether the assumption holds.","tokens_in":7723,"tokens_out":7319,"duration_ms":79710,"concrete_test":"Recompute NDI scores and GMM Tier labels for each subject's own connectome (same edge thresholds and weightings as Section 2.2), then quantify agreement with the control-derived group Tier labels used in Section 2.4 using the adjusted Rand index, separately for ASD and control groups. If the ARI is low or differs substantially between groups, the group-to-individual label transfer is unreliable, and the reported Tier-specific group differences cannot be attributed to the NDI-defined important regions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference in Sections 2.4 and 4 is that Tier labels estimated from group-averaged connectomes (Section 2.2: edges present in at least 90% of subjects, weights averaged) can be transferred to individual connectomes, and that differences in transitivity and efficiency within Tier 1 and Tier 2 correspond to differences in the most important nodes for each individual. This transfer is not tested. The stability results in Section 3 compare NDI rankings across cohort, patient, and control group connectomes; they establish only that group-average NDI rankings are similar, not that the resulting Tier boundaries match the NDI-defined topology of any individual subject. For functional connectomes, individual topology is noisy, and for ASD/ABIDE, site and diagnostic heterogeneity make systematic deviations plausible. If a node assigned to Tier 1 on the control group connectome is not high-NDI for many ASD subjects, or vice versa, then the per-Tier network measures are computed on different effective node sets, and a between-group difference can be produced by label misassignment rather than by true differences in the topology of important regions. The Discussion explicitly defers individual-level NDI estimation to future work, which is an acknowledgment that the required validation is missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies the Network Dependency Index (NDI) subnetwork framework to resting-state functional connectomes from 819 ABIDE subjects (440 controls, 379 ASD), parcellated into 90 AAL regions. Group-averaged connectomes are constructed for the full cohort, control-only, and patient-only groups under three edge thresholds (0.01, 0.03, 0.05) and three weighting schemes (positive, negative, absolute); NDI scores are computed on each group connectome and Tier labels are assigned via a 3-component Gaussian mixture model on log-transformed NDI values. The authors compare NDI rankings across group connectomes with rank-biased overlap (RBO), report that rankings are highly stable, then apply control-derived Tier labels to individual connectomes to compare transitivity and efficiency between ASD and controls within each Tier. Significant group differences are reported in Tier 1 for absolute weights at all thresholds and in Tiers 1/2 for positive weights at selected thresholds. The conclusion is that NDI can be applied to functional connectomes, produces stable subnetworks across group compositions, and can reveal ASD-related differences concentrated in the most important nodes.","tokens_in":7961,"tokens_out":7742,"duration_ms":78682,"significance":"If the central claims were fully substantiated, the paper would extend a promising, largely parameter-free subnetwork stratification method from structural to functional connectomics and provide a practical tool for characterizing distributed disease effects. The study has notable strengths: a large multi-site ABIDE sample, systematic exploration of thresholds and weighting schemes, transparent reporting of a failure mode (negative-only weights), and use of a rank-based stability metric. However, the current analysis does not yet support the application claim: the ASD-control comparison relies on unadjusted Mann-Whitney tests across many combinations, the transfer of group-derived Tier labels to individuals is not validated, and potential ABIDE site effects are not controlled. These issues are fixable within the manuscript's scope, so the work is a plausible methods contribution rather than a definitive clinical finding.","major_comments":[{"comment":"The ASD-versus-control comparisons consist of Mann-Whitney tests on transitivity and efficiency for four Tiers, three thresholds, and two retained weighting schemes, i.e., roughly 48 tests, with significance set at p<0.05 and no correction for multiple comparisons. The sparse pattern of significant results across Tiers and thresholds is consistent with what would be expected under the null, and no effect sizes or confidence intervals are reported. The conclusion that NDI Tiers \"can be utilized to show group differences\" is therefore not yet supported. Please apply an FDR or permutation-based correction across the full set of tests, pre-specify a primary Tier and measure, and report effect sizes.","section":"Section 2.5 and Figures 2-3"},{"comment":"Tier labels are estimated on group-averaged connectomes (edges present in at least 90% of subjects, weights averaged) and then transferred to each individual's connectome to compute per-Tier transitivity and efficiency. The paper does not check whether the group Tier labels correspond to the high-NDI nodes of individual subjects; the RBO stability results in Table 2 concern only the rank ordering of the three group connectomes, not the match between group labels and individual topology. The Discussion explicitly defers individual-level NDI estimation to future work, which is an acknowledgment that this load-bearing assumption is unvalidated. If the NDI topology of ASD subjects deviates from the control-derived labels, the reported per-Tier group differences could be produced by mislabeled node sets rather than by true differences in the topology of important regions. Please validate the transfer, for example by computing individual NDI-based Tier assignments and reporting overlap with the group labels, or substantially soften the conclusion.","section":"Section 2.4 and Discussion"},{"comment":"The ASD-control comparisons do not account for ABIDE site, age, sex, head motion, or other confounders. ABIDE is a multi-site repository with well-documented site effects, and the cohort spans a wide age range (mean 16.4 years, SD 7.1), so the observed differences in Tier-specific measures could reflect site or motion differences rather than diagnosis. Please include site as a covariate (e.g., mixed-effects modeling or stratified within-site analysis) or demonstrate that the results are consistent across sites before claiming diagnostic group differences.","section":"Section 3 and Table 1"},{"comment":"The stability claim \"irrespective of the group connectome\" is supported only for NDI rankings via RBO, not for the Tier stratification that is actually used in the subsequent analysis. Tier boundaries are determined by the fitted GMM centers, and two connectomes with nearly identical rankings can still yield different GMM fits and hence different node-to-Tier assignments. The paper does not report the number of nodes per Tier or the overlap of Tier assignments across the cohort, control, and patient connectomes. Without these values, the conclusion that the subnetworks themselves are stable across populations is overstated; please report them or restrict the claim to rank stability.","section":"Sections 2.2-2.3 and Table 2"}],"minor_comments":[{"comment":"There is a typo \"afer\" (should be \"after\") in the sentence \"most subnetwork stratification, afer the brain network has been estimated,\" and the phrase \"with respects to\" should be \"with respect to\".","section":"Section 1"},{"comment":"The text describes NDI as requiring \"no user-parameter,\" but the framework fixes the number of Gaussian components at 3 and the results depend on the chosen edge threshold and weighting scheme. Please rephrase to something like \"no user-selected subnetwork parameter\" or explicitly acknowledge the remaining choices.","section":"Abstract and Section 1"},{"comment":"It is unclear whether transitivity and efficiency are computed on the induced subgraph of each Tier or on the full network with the node set restricted to that Tier; please state the exact definition used.","section":"Section 2.4"},{"comment":"For the negative-weight networks, the paper reports that no Gaussian fit was possible but does not characterize why; reporting the range or distribution of NDI values for the neg networks would help readers judge whether the failure is technical or substantive.","section":"Section 3"},{"comment":"The RBO matrix is hard to read because the same value appears twice and the row and column labels mix threshold and weighting; consider presenting a single triangular matrix with combined labels such as \"abs-0.01\" and \"pos-0.01\".","section":"Table 2"},{"comment":"The region names \"cuneous\" and \"precuneous\" should be \"cuneus\" and \"precuneus,\" and \"Mann-Whitney-Wilcox test\" is more commonly called the Wilcoxon rank-sum test.","section":"Results text"},{"comment":"The manuscript does not state software versions (e.g., nilearn) or provide analysis code; a code/data availability statement would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a feasibility and stability demonstration for NDI on functional connectomes. The main obstacles to acceptance are the uncorrected multiple testing, the unvalidated group-to-individual label transfer, and the absence of site controls; these are addressable in a revision. If the authors prefer to keep the current scope, they should substantially temper the ASD-difference conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe short version: this is a modest but honest extension of the authors' own NDI subnetwork framework from structural to functional connectomes, using ABIDE. The novel bits are the stability analysis across group connectomes and the exploratory ASD vs. control comparisons. It is not a breakthrough, but it does what it claims, with caveats.\n\nWhat is new and what works: NDI itself is from Schirmer et al. 2018. Here they show it works on 90-node rsfMRI connectomes, and that NDI rankings are nearly identical whether the group connectome comes from the whole cohort, controls only, or patients only (RBO differences <0.005). That is a real, useful descriptive finding. They also test three edge thresholds and three weighting schemes, which is thorough for a short paper. The writing is clear, and they flag the main limitations themselves, including the fact that they used group-derived Tier labels rather than estimating individual-level NDI, which they defer to future work.\n\nThe soft spots are statistical and conceptual. First, the ASD-control differences: they run Mann-Whitney tests over a large number of combinations (roughly 2 weightings x 3 thresholds x 4 tiers x 2 measures) with no multiple-comparison correction, no effect sizes, and no modeling of ABIDE site effects. Some findings are at p<0.05 only, so I would treat the specific significant Tiers as hypothesis-generating, not a confirmed result. Second, the stress-test concern is real: applying group-level Tier labels to individual connectomes assumes those labels match each individual's own NDI-defined topology. They never validate that assumption, and they admit it by listing individual-level NDI as future work. In a heterogeneous dataset like ABIDE, that could produce spurious group differences if high-NDI nodes in controls are not high-NDI in ASD. Third, the exclusion of negative-only weights because the GMM failed, and the exclusion of Tier 4 at one threshold, are post hoc decisions; not fatal, but they should be reported more transparently as limitations.\n\nWho is this for? People working on subnetwork stratification in brain connectomics, especially those applying NDI or rich-club-like approaches. It is not a methodological breakthrough, but the stability result is a useful addition, and the autism findings are worth a follow-up with proper statistics. The self-citation is fine here; the framework is theirs and the extension is genuine.\n\nRecommendation: I would accept it for peer review — it is exactly the kind of modest, clearly-scoped application that a workshop or short paper venue should consider — and I would ask the authors to revise with corrected statistics and ideally some individual-level validation. Without that, the stable findings stand, but the disease-difference claim stays weakly supported.","headline":"A modest, honest application of the authors' own NDI framework to functional connectomes; the stability results hold up, but the ASD-difference claims need stronger statistics and individual-level validation.","tokens_in":8454,"tokens_out":2291,"would_cite":false,"duration_ms":23144,"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":"The network dependency index, a parameter-free measure of a node's contribution to network efficiency, extends to resting-state functional connectomes and places autism-related differences in high-importance subnetworks.","keywords":["network dependency index","subnetwork stratification","resting-state fMRI","functional connectome","autism spectrum disorder","network efficiency","transitivity","Gaussian mixture model"],"falsifier":"Recompute NDI Tiers from each individual's own connectome and compare them with the group-derived labels; if subject-level Tier assignments disagree widely with group labels, or if the ASD-control differences in Tier 1 transitivity and efficiency disappear when Tiers are defined per subject, the central conclusion fails.","tokens_in":7540,"feed_emoji":"🧠","tokens_out":10201,"duration_ms":89256,"temperature":0.7,"pith_summary":"The paper extends the network dependency index (NDI)—a score for how much removing a node lowers a network's overall efficiency—from structural brain networks to resting-state functional connectomes. It asks whether the NDI-based Tier subnetwork labels stay stable when the group connectome used to define them is built from the whole cohort, from patients only, or from controls only, and whether those labels can expose autism-related topology. The answer offered is yes on both counts: Tier assignments agree almost completely across the three group connectomes (ranked-biased overlap differences below 0.005), and rank-based tests find significant differences between individuals with ASD and controls in transitivity and/or efficiency in Tier 1, the subnetwork of most important nodes, with positive-weight analyses also implicating Tier 2 at some thresholds. This matters because NDI stratifies the brain into subnetworks without a user-set parameter, giving connectomics a more reproducible way to localize disease effects in hub-like regions.","feed_headline":"Autism differences land in the brain's hub tiers","feed_subtitle":"Stable tiers rank node importance; autism-control differences concentrate in the highest tiers.","key_machinery":"The load-bearing object is the network dependency index (NDI), a per-node score defined as the mean loss in network efficiency—measured through information $I_{ij}=1/D_{ij}$, where $D$ comes from shortest-path distances on inverse connection weights—when that node is removed. Nodes with NDI zero form Tier 4; the remaining log-transformed scores are split into three further Tiers by a Gaussian mixture model with three components, with subnetwork boundaries at halfway points between centers. Tier labels are defined once on a group-averaged connectome (edges present in at least 90% of subjects, with weights averaged) and then transferred to each individual connectome, where Tier-specific transitivity and global efficiency are compared between groups. Stability is assessed with ranked-biased overlap, a similarity measure that weights agreement at the top of the importance ranking more heavily, and group differences are tested with rank-based tests.","core_discovery":"The central discovery is that the NDI subnetwork framework, originally validated on structural connectomes with 170 regions, transfers to functional connectomes with 90 regions and remains stable when the population used to construct the group connectome changes. On resting-state fMRI data from 819 participants (440 controls, 379 individuals with ASD), NDI scores and the Tier labels derived from a three-component Gaussian mixture model agreed across cohort, patient-only, and control-only group connectomes for every threshold and weighting combination tested; the choice of absolute versus positive edge weights moved node rankings more than the edge threshold (0.01, 0.03, or 0.05). When Tier labels from the control connectome were applied to individual connectomes, statistically significant ASD-control differences appeared in Tier 1 for absolute weights at all thresholds and in Tier 1 or Tier 2 for positive weights, supporting the claim that autism-related functional differences concentrate in nodes of highest NDI importance. The right median cingulate and paracingulate gyri was the only region identified across all thresholds and weighting schemes.","pith_inferences":["Editorial inference: the paper's transfer of group-level Tier labels to individuals assumes those labels represent each individual's functional organization; recomputing NDI per subject and measuring agreement with group labels would directly test this assumption, especially across the multi-site ASD sample.","Editorial inference: because NDI needs no user-set parameter, the Tier framework could serve as a data-driven prior for region-of-interest analyses in other disorders, but only after validation across atlases and independent cohorts.","Editorial inference: the paper compares only group-level labels; an individual-level strategy could reveal ASD subtypes with different hub profiles, a direction the authors explicitly leave to future work."],"forward_implications":["NDI subnetwork stratification is not limited to high-resolution structural connectomes: it works on resting-state functional connectomes parcellated into 90 regions, roughly half the resolution of the original 170-region structural study.","Subnetwork labels are effectively invariant to whether the group connectome is built from the whole cohort, from patients only, or from controls only, with ranked-biased overlap differences below 0.005 across conditions.","Edge threshold level changes NDI node rankings less than the choice of absolute versus positive edge weights, so threshold choice is a secondary source of variation.","Applied to individuals, the group-defined tiers place ASD-control differences in transitivity and efficiency preferentially in Tier 1 (and Tier 2 for positive weights), localizing disease-related topology to the regions of highest NDI importance.","The right median cingulate and paracingulate gyri emerges as the only region consistently in the significant tiers across all thresholds and weighting schemes, giving a candidate anatomical focus for follow-up."],"supporting_citations":[{"why":"Defines the network dependency index and the Tier subnetwork framework that this paper extends from structural to functional connectomes.","marker":"[19]"},{"why":"Supplies the multi-site resting-state fMRI cohort of individuals with autism and typically developing controls.","marker":"[9]"},{"why":"Defines transitivity and global efficiency, the two network measures used to compare Tiers between groups.","marker":"[15]"},{"why":"Supplies ranked-biased overlap, the similarity measure used to show NDI node rankings are stable across group connectomes.","marker":"[23]"},{"why":"Describes the covariance-based construction of functional edge weights that produces the connectomes analyzed in the paper.","marker":"[22]"}],"fun_headline_variants":["Autism hubs stand out in brain network tiers","Brain hub tiers flag autism connectivity differences","NDI tiers pinpoint autism-linked brain hubs","Autism differences cluster in high-importance brain tiers","Stable brain tiers reveal autism hub shifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Tier labels computed on group-averaged connectomes describe each individual's functional organization; if individual connectomes—especially in a heterogeneous ASD cohort collected across multiple sites—organize differently, group-level labels can misclassify regions and the reported group differences may be artifacts of mislabeling.","fun_headline_variants_meta":{"raw":{"variants":["Autism hubs stand out in brain network tiers","Brain hub tiers flag autism connectivity differences","NDI tiers pinpoint autism-linked brain hubs","Autism differences cluster in high-importance brain tiers","Stable brain tiers reveal autism hub shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00036,"raw_usage":{"total_tokens":1986,"prompt_tokens":1026,"completion_tokens":960,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":892}},"tokens_in":642,"tokens_out":960,"duration_ms":7682,"temperature":1.0,"reasoning_tokens":892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:20:55.646307+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute NDI Tiers from each individual's own connectome and compare them with the group-derived labels; if subject-level Tier assignments disagree widely with group labels, or if the ASD-control differences in Tier 1 transitivity and efficiency disappear when Tiers are defined per subject, the central conclusion fails.","supporting_citations":[{"cited_title":"ACM Transactions on Information Systems (TOIS) 28(4), 20 (2010)","cited_arxiv_id":null,"evidence_quote":"Supplies ranked-biased overlap, the similarity measure used to show NDI node rankings are stable across group connectomes."}],"review_version":1}