{"id":"427b7dbb-a3cd-4fb1-a357-5f1f76dbd516","arxiv_id":"2608.07393","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A federated framework that decomposes multi-site brain dynamic connectivity into disease, phenotype, and scanner parts and aligns site prototypes to improve ASD and ADHD classification.","lead":"FedDOSE is a federated learning method that splits brain-scan connectivity data from different hospitals into disease, patient, and scanner parts, so models can learn disease patterns without pooling raw data. It reports higher accuracy than seven existing federated methods on autism and ADHD diagnosis across three multi-site fMRI datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Softmax-gated masks in Eq. 5 cannot by construction produce orthogonal disease/phenotype/scanner subspaces; without a direct separation metric, the paper's central decomposition rationale is not established.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the claimed decomposition of site effects into orthogonal disease, phenotype, and scanner subspaces is asserted without proof or direct measurement. My reading sharpens this into a concrete architectural objection: softmax gating in Eq. 5 produces strictly positive masks, so the masked tensors overlap and cannot partition the core-tensor variance. This is an internal-consistency issue, not a debate against an external consensus. It does not overturn the reported empirical comparisons, which may still hold because the added losses and alignment steps could improve accuracy even if the subspace story is not literally true. It does, however, undermine the paper's central explanatory claim, which is what the reader already flagged. Since the reader's CONDITIONAL verdict already requires addressable revision rather than acceptance, my concern does not move the verdict; it strengthens the reason for the condition. A direct measurement of mask overlap and cross-embedding correlation would settle whether the mechanism claim survives.","tokens_in":14356,"tokens_out":5848,"duration_ms":64311,"concrete_test":"Retrain FedDOSE on ABIDE-I with the reported defaults. After convergence, measure subspace separation on held-out subjects: (1) compute the elementwise overlap between mask pairs, e.g. mean(|vec(Md) ⊙ vec(Mb)|) / (||Md||_F ||Mb||_F), and similarly for (Md, Mv) and (Mb, Mv); (2) compute the mean absolute canonical correlations (or cosine similarities) between the learned embeddings zd, zb, and zv. If the mask overlaps are substantially above zero or the cross-embedding correlations are non-negligible, the 'orthogonal subspace' claim in Section 3.3 is not supported by the architecture, and the paper must either add an explicit orthogonalization constraint or re-state the claimed mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central explanatory claim is that FedDOSE's accuracy gains stem from explicitly decomposing site effects into disease, phenotype, and scanner subspaces. This requires that the three subspaces actually separate site-related variance. The construction in Eq. 5, [Md, Mb, Mv] = Softmax(MLP(sum_m G^m)), does not enforce or imply this. A softmax over the three mask channels assigns a strictly positive weight to every element in all three masks, so the masked tensors Md⊙G, Mb⊙G, and Mv⊙G overlap; they are not a partition of the core-tensor variance and are not orthogonal. No orthogonality or identifiability constraint is imposed on the masks or on the resulting embeddings zd, zb, zv. The orthonormality penalty in L_recon (Eq. 4) applies to the spatial factors U_s^m only, not to the subspace decomposition. The gradient-reversal terms in Eq. 6 encourage zd and zb to be poor predictors of scanner labels, but they do not guarantee that disease-relevant signal survives, nor that the three embedding spaces are independent. Therefore the empirical gains in Tables 1-4 and the ablations demonstrate that the added losses help, but they do not establish the stated mechanism: that site differences are decomposed into separable disease, phenotype, and scanner subspaces. This is a load-bearing gap because the paper's novelty claim rests on this decomposition, not merely on the aggregate accuracy numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedDOSE, a federated learning framework for multi-site resting-state fMRI analysis that models dynamic functional connectivity (dFC). The framework combines a Modularity-Guided Tucker Decomposition (MGTKD) block for dimensionality reduction, a softmax-gated decomposition of the aggregated core tensor into disease, phenotype, and scanner subspaces, and a prototype-alignment mechanism based on optimal transport barycenters and Procrustes analysis. The authors evaluate FedDOSE on three public datasets (ABIDE-I, ABIDE-II, ADHD-200) using three selected sites per dataset, reporting site and global accuracy improvements over several federated learning baselines, plus ablations, sensitivity analyses, and a biomarker illustration. The central claim is that FedDOSE outperforms existing FL methods in both site and global accuracy and that this improvement stems from explicitly decomposing site differences into separable disease, phenotype, and scanner subspaces.","tokens_in":14793,"tokens_out":5459,"duration_ms":64149,"significance":"If the proposed decomposition mechanism were rigorously established, FedDOSE would be a meaningful contribution to privacy-preserving multi-site fMRI analysis: it targets dynamic connectivity rather than static FC, introduces a modular tensor-decomposition block that leverages known brain architecture, and combines prototype learning with OT-based alignment to handle site heterogeneity. The paper includes a reasonably broad set of baselines, per-site results, ablations, a cross-site generalization study, and a sensitivity analysis, which are useful for assessing the empirical behavior of the framework. However, the paper does not provide code, uses a post hoc subset of sites, and the central explanatory claim—that the accuracy gains arise from decomposing site effects into orthogonal subspaces—is not supported by any direct measurement or identifiability argument. The empirical results are internally consistent and suggest that the added losses help, but they do not yet substantiate the proposed mechanism as the cause of the observed gains.","major_comments":[{"comment":"The paper states that softmax-based gating 'partitions the variance in the core tensors into three orthogonal subspaces,' but the construction in Eq. (5) does not enforce or imply orthogonality. A softmax over the three mask channels produces strictly positive weights, so the masked tensors M_d⊙G, M_b⊙G, and M_v⊙G overlap and are not a partition of the core-tensor variance. No orthogonality or identifiability constraint is imposed on the masks or on the resulting embeddings z_d, z_b, z_v; the orthonormality penalty in Eq. (4) applies only to the spatial factors U_s^m. Because the subspace-decomposition rationale is the paper's novelty claim, the authors should provide direct evidence of separation—for example, cosine similarities between the embeddings, overlap measures of the masked tensors, or an ablation that uses hard one-hot masks—or temper the claim to state that the losses encourage separation rather than enforce it.","section":"§3.3, Eq. (5)"},{"comment":"The conclusion that 'FedDOSE outperforms other FL methods' is not supported by the reported significance tests with respect to the strongest baseline. In Tables 1, 2, and 4, none of the FedDOSE versus FedGMKD comparisons are marked with an asterisk, and several FedDOSE versus FedAli comparisons also lack asterisks. With only five seeds, the Mann-Whitney U test has very low power, so the absence of a significant difference does not establish superiority. The authors should report effect sizes with confidence intervals, use a paired test across seeds, or increase the number of seeds, and should frame the headline claim accordingly.","section":"§4.4, Tables 1–4"},{"comment":"The evaluation uses only three selected sites per dataset rather than the full multi-site data, and the selection is made post hoc by the authors (e.g., excluding NI for low sample size and KKI/WUSTL for severe class imbalance in ADHD-200). This selection is performed on the same datasets used for evaluation, which undermines the generalizability claim and can bias the comparison if excluded sites are more heterogeneous. Moreover, the abstract states 'approximately 2000 subjects,' but the selected cohort contains about 1,368 subjects (ABIDE-I: 323, ABIDE-II: 218, ADHD-200: 827). Please either report results on all available sites (or a pre-registered split), justify why the selection is not outcome-dependent, and correct the subject count in the abstract.","section":"§4.1, Datasets and Preprocessing"},{"comment":"The gradient-reversal scanner loss encourages z_d and z_b to be poor predictors of scanner labels, but it does not guarantee that disease-relevant signal is preserved. If scanner variation correlates with disease or phenotype (which is common in multi-site studies), the GRL can remove informative variance and degrade performance. The paper does not measure how much scanner information remains in z_d after training, nor how disease accuracy varies under scanner shift. The authors should add a quantitative analysis, such as scanner-label prediction accuracy from z_d before and after the GRL, or a stratified evaluation where each site's model is tested on the other sites (they do provide cross-site generalization in Table 5, but without isolating scanner effects). This is necessary to substantiate the claim that the decomposition separates scanner variance without erasing disease signal.","section":"§3.3, Eq. (6)"}],"minor_comments":[{"comment":"The abstract's initial description of the datasets is inaccurate: the paper uses a subset of each dataset, not the full ABIDE-I, ABIDE-II, and ADHD-200 cohorts, and the total number of subjects is about 1,368, not 'approximately 2000.' Please adjust the abstract to reflect the actual experimental cohort.","section":"Abstract"},{"comment":"The notation 'log(X)' is ambiguous for a tensor X∈R^{N×N×T}; it should specify that the matrix logarithm is applied per temporal slice X_t. Also, Eq. (1) uses ×_1 and ×_2 with B_m^T, which selects ROIs, but it is not clear whether the temporal dimension is indexed by the third mode; a brief clarification would help.","section":"§3.1, Eq. (1)"},{"comment":"The optimal transport barycenter in Eq. (11) is written for distributions p and p_k, but the prototypes p_k are matrices or vectors, not distributions. Please specify how the 2-Wasserstein distance is computed for these prototypes (e.g., by treating each prototype's columns as empirical samples or by defining a matrix distance).","section":"§3.5, Eq. (11)"},{"comment":"The text states that removing the phenotypic and scanner subspaces leads to a 'significant drop in performance,' but not all ablations are statistically significant relative to FedDOSE; for example, the w/o OT and w/o L_proto rows have no asterisks in several columns. Please qualify the strength of the ablation conclusions.","section":"§4.5, Table 3"},{"comment":"The sentence 'performance with respect to λ typically peaked at 0.2 before declining' is ambiguous because the analysis reports global accuracy only; please state whether the stated trends refer to global accuracy, site accuracy, or a specific site, and consider reporting the variability across seeds.","section":"§4.6, Sensitivity Analysis"},{"comment":"No code or implementation details are provided beyond the text. For a federated learning method, releasing code (or at least a pseudocode/pipeline description) would substantially improve reproducibility and allow others to verify the subspace-decomposition behavior.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and contains a substantial amount of experimental work, but the central mechanism claim (orthogonal site-effect decomposition) is not validated, and the empirical superiority claim is weakened by the lack of statistical significance against the strongest baseline. I recommend major revision with the explicit requirements that the authors either provide direct evidence of subspace separation or soften the mechanistic language, report results on all sites or justify the selection, and strengthen the statistical analysis. The paper would fit the journal if these issues are resolved; at present, the contribution is more incremental than the abstract suggests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FedDOSE is a useful, well-run empirical paper with a slightly oversold central story. The framework combines modularity-guided Tucker decomposition of dFC tensors with site-level decomposition losses and OT/Procrustes prototype alignment. That combination is new in the FL-fMRI literature, and the experiments support it: FedDOSE beats seven baselines on three datasets, and the ablations show each component contributes. I'd send this to a serious referee.\n\nWhere the paper wobbles is the mechanism. Section 3.3 says the softmax-gated masks 'partition' the core-tensor variance into three orthogonal subspaces. That is not what a softmax over three mask channels does: every element gets positive weight in all three masks, so the masked tensors overlap and nothing enforces orthogonality. The gradient-reversal losses encourage the embeddings to be poor predictors of scanner labels, but that is a soft constraint, not a demonstration that disease, phenotype, and scanner variance actually separate. The ablations show removing these losses hurts accuracy, which is good evidence they matter, but it doesn't confirm the stated decomposition. I'd want a direct separation metric or a variant with hard masks before accepting the interpretability story. This is a real soft spot, but not a fatal one: the empirical gains stand on their own.\n\nOther issues are smaller. The abstract says 'approximately 2000 subjects,' but after site selection the cohort is about 1370. That should be corrected. The site choices are reasonable (largest, balanced) but they narrow the multi-site claim. No code is released, which hurts reproducibility for a methods paper. Baseline tuning is described as 'default hyperparameters,' so some baselines may be at a disadvantage; worth asking about in review.\n\nThe citation pattern looks fine, and there's no smoke-and-mirrors in the derivations. The method is sensible, the evaluation is thorough, and the limitations section honestly acknowledges computational overhead. For researchers working on FL for fMRI, this is worth engaging with. My recommendation: accept for peer review, but ask for a clearer demonstration of subspace separation, corrected subject counts, and code release.","headline":"Solid FL-for-fMRI paper whose accuracy gains are credible, but the 'orthogonal subspace decomposition' mechanism is asserted more strongly than it is demonstrated.","tokens_in":15190,"tokens_out":2457,"would_cite":true,"duration_ms":24145,"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":"FedDOSE, a federated learning framework for dynamic brain connectivity, splits site differences into disease, phenotype, and scanner subspaces and outperforms seven federated baselines on three multi-site fMRI datasets.","keywords":["federated learning","dynamic functional connectivity","site effect decomposition","Tucker decomposition","optimal transport","Procrustes analysis","autism spectrum disorder","ADHD"],"falsifier":"Train a scanner classifier on the disease-space embedding z_d after FedDOSE training; if its accuracy is significantly above chance, or if the Frobenius norm of the product of the disease and scanner masks M_d ⊙ M_v is large relative to the individual masks, the claimed subspace separation has not been achieved.","tokens_in":14141,"feed_emoji":"🧠","tokens_out":4859,"duration_ms":44131,"temperature":0.7,"pith_summary":"The paper claims that federated models trained on multi-site fMRI struggle not merely because of distribution shift but because site differences remain entangled, and that explicitly decomposing site variability into disease, phenotype, and scanner subspaces, then aligning class prototypes across sites, yields a privacy-preserving model that approaches centralized accuracy. The framework, FedDOSE, works on dynamic functional connectivity (dFC), which changes on the order of seconds, rather than the static connectivity most federated fMRI methods use. On ABIDE-I, ABIDE-II, and ADHD-200, FedDOSE beats seven federated baselines on both site-level and global accuracy, with the largest gains on the most heterogeneous datasets, and comes within roughly 2 percentage points of a centralized upper bound on ABIDE-I. If the claim holds, multi-site psychiatric imaging studies can pool data across institutions without sharing raw scans and with only a small accuracy penalty.","feed_headline":"FedDOSE beats seven federated baselines on ASD and ADHD scans","feed_subtitle":"Separating disease, phenotype, and scanner signals gets multi-site accuracy within 2 percent of centralized training.","key_machinery":"The core structural elements are the Modularity-Guided Tucker Decomposition (MGTKD) block, which compresses each subject's high-dimensional dFC tensor into per-module core tensors and shared spatial and temporal factors while preserving brain modular structure; the softmax-gated masks of Equation (5), which partition the aggregated core-tensor variance into three subspaces; and the two-stage global aggregation in which an optimal-transport barycenter (Equation 11) defines a reference prototype distribution and Procrustes analysis (Equation 12) rotates each site's prototypes into that shared reference before weighted aggregation. The MGTKD block supplies compressed, interpretable features; the gating supplies the claimed separation of site effects; and the OT-Procrustes alignment corrects for the misalignment of latent components across independently decomposed sites.","core_discovery":"The central claim is that explicit, supervised decomposition of site effects into three learned subspaces—disease, phenotype, and scanner—carried by softmax-gated masks on a modular Tucker decomposition of dynamic functional connectivity tensors, combined with optimal-transport barycenter and Procrustes alignment of class prototypes, lets federated models nearly match centralized performance. The authors establish this through experiments on three multi-site datasets, showing that FedDOSE consistently achieves the highest site and global accuracy across all sites, with an average improvement of about 4 percent over the strongest baseline on ABIDE-I and ABIDE-II and about 3 percent on ADHD-200. Cross-site generalization accuracy also exceeds that of all baselines, and the ablation study shows that removing the phenotype and scanner losses, the Procrustes alignment, the optimal-transport reference, or the prototype regularization each degrades performance, indicating that the decomposition and alignment are what drive the gains.","pith_inferences":["The orthogonality of the three subspaces is asserted but not measured; one could directly test whether the disease and scanner masks overlap substantially, which would predict partial confounding in the disease embedding.","If the decomposition works as intended, the same framework could be ported to other multi-site neuroimaging modalities or to harmonization tasks, replacing post-hoc batch correction with a supervised, task-aligned decomposition that is learned end-to-end.","Because the global prototype is built from aligned spatio-temporal coupling maps, FedDOSE might serve not only as a classifier but also as a method for cross-site discovery of dynamic connectivity biomarkers, although the paper does not pursue this direction.","The reported gains of roughly 2 to 4 percent, while consistent, are modest; the strongest test of the decomposition rationale would be whether the advantage persists under larger scanner variability or with noisier disease labels."],"forward_implications":["FedDOSE narrows or closes the gap to a centralized, privacy-violating upper bound on site accuracy across all three datasets, making federated learning a viable route for multi-site rs-fMRI consortia.","The explicit decomposition of site differences, rather than a generic domatin-adversarial or harmonization step, is what drives the accuracy gains, as shown by the ablation results on ABIDE-I.","Because the framework operates on dynamic connectivity, it captures temporal connectivity fluctuations that static-connectivity federated methods miss, which may improve detection of neurodevelopmental disorders.","Cross-site generalization accuracy is highest for FedDOSE, indicating that the learned representations capture invariant disease features rather than site-specific memorization.","The modular biomarker analysis, based on outer products of spatial factors, flags connections involving sensorimotor and dorsal attention networks that align with existing ASD literature, suggesting a path toward interpretable federated biomarkers."],"supporting_citations":[{"why":"Provides the standard Federated Averaging scheme that FedDOSE uses to aggregate local model parameters.","marker":"[22]"},{"why":"Supplies the Tucker decomposition method on which the Modularity-Guided Tucker Decomposition block is built.","marker":"[33]"},{"why":"Supplies the optimal-transport barycenter formulation used to construct the reference prototype distribution for cross-site alignment.","marker":"[16]"},{"why":"Supplies the generalized Procrustes analysis used to rotate site prototypes into the shared reference space.","marker":"[10]"},{"why":"Supplies the gradient reversal layer used to enforce scanner invariance in the disease and phenotype subspaces.","marker":"[8]"},{"why":"Supplies the ABIDE-I dataset used for the main ASD detection experiments and ablation studies.","marker":"[5]"},{"why":"Supplies the ABIDE-II dataset used to validate the framework on a second ASD cohort.","marker":"[4]"},{"why":"Supplies the ADHD-200 dataset used to demonstrate generalization to a different neurodevelopmental disorder.","marker":"[2]"},{"why":"Serves as the strongest prototype-based federated baseline that FedDOSE is compared against.","marker":"[43]"},{"why":"Serves as a prototype-exchange federated baseline that the framework must outperform to establish its alignment strategy.","marker":"[32]"}],"fun_headline_variants":["FedDOSE splits site effects to reach near-central accuracy in federated brain scans","Federated MRI model untangles scanner effects to match centralized training","FedDOSE: separating disease, phenotype, and scanner signals in federated brain nets","Federated brain model hits within 2% of centralized by decomposing site effects","Site-effect decomposition lets FedDOSE match centralized federated brain scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes the softmax-gated masks partition the aggregated core tensor into truly orthogonal disease, phenotype, and scanner subspaces, and that the gradient-reversal scanner losses erase scanner variance without removing disease-relevant signal.","fun_headline_variants_meta":{"raw":{"variants":["FedDOSE splits site effects to reach near-central accuracy in federated brain scans","Federated MRI model untangles scanner effects to match centralized training","FedDOSE: separating disease, phenotype, and scanner signals in federated brain nets","Federated brain model hits within 2% of centralized by decomposing site effects","Site-effect decomposition lets FedDOSE match centralized federated brain scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001052,"raw_usage":{"total_tokens":4435,"prompt_tokens":979,"completion_tokens":3456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":3354}},"tokens_in":595,"tokens_out":3456,"duration_ms":22487,"temperature":1.0,"reasoning_tokens":3354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:12:09.310695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a scanner classifier on the disease-space embedding z_d after FedDOSE training; if its accuracy is significantly above chance, or if the Frobenius norm of the product of the disease and scanner masks M_d ⊙ M_v is large relative to the individual masks, the claimed subspace separation has not been achieved.","supporting_citations":[{"cited_title":"Energy-guided continuous entropic barycenter estimation for general costs.Advances in Neural Information Processing Systems, 37:107513–107546, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal-transport barycenter formulation used to construct the reference prototype distribution for cross-site alignment."},{"cited_title":"Generalized procrustes analysis.Psychometrika, 40(1):33–51, 1975","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized Procrustes analysis used to rotate site prototypes into the shared reference space."},{"cited_title":"Enhancing studies of the connectome in autism using the autism brain imaging data exchange ii.Scientific data, 4(1):1–15, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the ABIDE-II dataset used to validate the framework on a second ASD cohort."},{"cited_title":"Adhd-200 global compe- tition: diagnosing adhd using personal characteristic data can outperform resting state fmri measurements.Frontiers in systems neuroscience, 6:69, 2012","cited_arxiv_id":null,"evidence_quote":"Supplies the ADHD-200 dataset used to demonstrate generalization to a different neurodevelopmental disorder."},{"cited_title":"Fedgmkd: An efficient prototype feder- ated learning framework through knowledge distillation and discrepancy-aware aggregation","cited_arxiv_id":null,"evidence_quote":"Serves as the strongest prototype-based federated baseline that FedDOSE is compared against."}],"review_version":1}