{"id":"3b40c001-236e-4975-85ce-75013c30080b","arxiv_id":"2502.01885","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"DAFed combines domain-adversarial training, feature disentanglement, and contrastive learning in a federated GCN framework to classify ASD and MCI from multi-site fMRI, claiming accuracy gains over existing federated baselines.","lead":"This paper introduces DAFed, a federated learning framework that trains a brain-connectivity classifier across multiple medical centers without sharing raw scans, by separating shared from site-specific features. It reports higher accuracy on autism and Alzheimer's-related datasets than several existing federated methods, along with maps of brain regions tied to each condition.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 is not executable as written: local sites are asked to backpropagate through Ltotal_source, a loss computed from central-site data they never receive, so the reported accuracies cannot be traced to the described federated procedure.","rationale":"I read the paper in good faith: the goal is a privacy-preserving federated framework for multi-site fMRI classification, and the central claim is state-of-the-art accuracy on ABIDE and ADNI-3. The reader's weakest assumption is precisely the one I would identify. The federated update rule is the backbone of the method; if it cannot be computed, all downstream results are disconnected from the proposed algorithm. Section II-C says the source loss is 'sent' to the target, but sending a scalar is insufficient for autograd-based backpropagation. The server could, in principle, send source gradients, and that would rescue executability, but that is not what the text describes, and it changes the privacy and communication model. This is an internal inconsistency in the algorithm specification, not a disagreement with community consensus. The empirical comparisons, source-site selection, and interpretability claims are secondary; none of them can validate a non-executable training loop. The paper includes no code release and no formal verification, so there is no independent check that the described algorithm was actually run. Since the reader already rejected on this ground, my verdict remains unchanged. I agree with the reader's assessment; no change to the verdict is needed.","tokens_in":19861,"tokens_out":3353,"duration_ms":34576,"concrete_test":"Implement Algorithm 1 exactly as written with one source client holding ABIDE-USM and one target client holding ABIDE-NYU, in PyTorch. At the target client, after receiving Θglobal and the scalar Ltotal_source, call loss.backward() on L = Ltotal_source + Ltotal_target_k. Record the error: Autograd will raise a RuntimeError because Ltotal_source is not connected to the target parameters. To determine whether the reported numbers depend on a hidden gradient channel, modify the protocol so the server transmits ∂Ltotal_source/∂Θglobal to each client, add it to the local gradient, and rerun Table 4. If accuracies change materially from 0.764 and 0.753, the original results rely on an unspecified mechanism; if they do not, the paper still needs a corrected protocol description.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DAFed achieves state-of-the-art accuracy through a privacy-preserving federated training procedure. The load-bearing assumption is that every local site can compute the gradient of the total objective, including Ltotal_source, with respect to its own parameters. In the stated protocol (Section II-C steps 3-5; Algorithm 1 lines 5-7), the central site broadcasts parameters and the scalar loss Ltotal_source, not source data or source gradients. A scalar loss value is a number; it carries no computational graph linking source samples to client parameters. Therefore ∂(Ltotal_source)/∂Θ_k is undefined at the client. The text says to use the total objective loss L = Ltotal_source + Ltotal_target to perform backpropagation and update Θ_k, but backpropagating through Ltotal_source requires either source data at every client or transmission of per-parameter source gradients. The paper provides neither. This is not a tunable weakness: as written, line 7 of Algorithm 1 cannot execute in any standard deep learning framework without access to source data or an explicit gradient message. Consequently, the empirical results in Tables 4 and 5 are not reproducible from the described method. Secondary concerns about source-site selection based on baseline performance and hyperparameter tuning on the test data are also present, but they would matter only if the algorithm were actually implementable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DAFed, a federated deep-learning framework for multi-site brain functional connectivity analysis. The method combines graph-convolutional feature extraction, feature disentanglement into domain-invariant and domain-specific components, multi-head attention integration, domain-adversarial training between a central labeled site and unlabeled local sites, contrastive learning on the domain-invariant representation, and an improved Score-CAM interpretation module. It is evaluated on ABIDE for ASD classification (four sites, 370 subjects) and on ADNI-3 for MCI classification (three scanners, 844 subjects). The authors report average accuracies of 0.764 (DAFed_L) and 0.753 (DAFed_U) on ABIDE, and 0.718 (DAFed_L) on ADNI-3, claiming superior performance over non-federated and federated baselines, and present ROI and functional-connectivity analyses.","tokens_in":20154,"tokens_out":5308,"duration_ms":53849,"significance":"If the described algorithm were executable and the evaluations were valid, the contribution would be practically important: it targets non-IID data in multi-site neuroimaging without sharing raw data, and it includes an interpretability module for biomarker discovery. The combination of federated learning, disentanglement, and adversarial domain adaptation is a sensible design, and the paper evaluates on two public cohorts with several baselines. The central algorithmic description, however, contains a gradient-computation error that prevents the procedure from running as written, and the evaluation protocol has outcome-dependent source selection and non-nested hyperparameter optimization. These issues are load-bearing for the headline accuracy claims, so the manuscript in its current form cannot support its conclusions.","major_comments":[{"comment":"The local update is not computable as written. The protocol broadcasts only the global parameters and the scalar loss Ltotal_source; a scalar loss carries no computational graph connecting source-domain samples to local parameters, so the derivative d(Ltotal_source)/dTheta_k is undefined at client k. Backpropagating through Ltotal_source would require either source data at every local site or an explicit per-parameter source-gradient message, and neither is described. Since line 7 of Algorithm 1 uses exactly this gradient, the algorithm cannot execute in any standard deep learning framework, and the accuracies in Tables 4 and 5 cannot be traced to the described federated procedure.","section":"Section II-C, Algorithm 1 lines 4-7"},{"comment":"The source domain is selected on the basis of the same baseline accuracies used in the comparison: the text states that USM is designated as the source because non-federated methods achieved better results on USM, and GE is chosen for the same reason on ADNI-3. Because 5-fold cross-validation is performed on the same data, this is outcome-dependent selection and can inflate the reported improvement of DAFed; it undermines the claim that DAFed outperforms all baselines. A nested or pre-registered source-selection rule, or an analysis showing robustness across all possible source choices, is needed.","section":"Results, ABIDE and ADNI-3 source selection"},{"comment":"Hyperparameters lambda1, lambda2, the learning rates, and the privacy noise level alpha are tuned with grid search on the same datasets used for the reported 5-fold cross-validation, and the cross-validation is not nested. This makes the reported accuracies optimistic and prevents a clean comparison with the baselines, whose hyperparameter choices are not reported with the same detail. The authors should either use a held-out validation fold for tuning or report sensitivity analyses over the hyperparameter ranges.","section":"Parameters setup and Experiments"},{"comment":"The privacy-preservation claim is not quantified. Adding Gaussian noise alpha*sigma to uploaded local models does not by itself constitute differential privacy; no privacy budget, composition analysis, or formal guarantee is provided, so the title's claim of being privacy-preserving is not supported. At minimum, the authors should state the precise privacy model and either provide a formal guarantee or soften the claim.","section":"Section II-D and Parameters setup"}],"minor_comments":[{"comment":"The sentence beginning 'Although methods such as ComBat can correct batch effects [10, 77]' is duplicated in the introduction and should be corrected.","section":"Introduction"},{"comment":"The notation 'mean(H) || max(H)' uses '||' for concatenation but the concatenation axis and the resulting dimension are not defined; please clarify.","section":"Section II-A, Eq. (2)"},{"comment":"The dimension flow from the GCN outputs to the first MLP in the disentangler is not explained: the input dimension 480 is not derived from the concatenated embeddings defined in Eq. (3).","section":"Table 3"},{"comment":"Line 9 averages Theta_k^t while line 7 updates Theta_k^{t+1}; the superscript/subscript indexing around the aggregation step is inconsistent and should be corrected.","section":"Algorithm 1"},{"comment":"The tables report only mean accuracy with standard deviation across folds; no paired significance tests or confidence intervals are given, so the claim that DAFed is 'best' is not statistically supported.","section":"Tables 4 and 5"},{"comment":"The manuscript does not include a code or data availability statement, which impedes reproducibility of the described method and experiments.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The manuscript should not be accepted in its current form. The Algorithm 1 issue is not cosmetic: it means the reported numbers cannot be produced by the described method, and the source-selection and hyperparameter-tuning issues further bias the evaluation. A corrected protocol would require explicitly transmitting source gradients or reformulating the optimization, and the experiments would need to be rerun under a cleaner evaluation design. The interpretability analysis could become a useful contribution if the model is made correctly executable and the evaluation bias is removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper combines domain-adversarial training, feature disentanglement, contrastive learning, and GCNs in a federated setting for multi-site fMRI classification. That specific integration is new, and the target problem—privacy-preserving collaborative learning across heterogeneous scanners—is real and worth working on. The paper also makes a genuine effort on interpretability with an improved Score-CAM adapted to GCNs, and the ablation study is a reasonable check that each module contributes. The soft spot is not minor: Algorithm 1 asks each local site to compute the gradient of Ltotal_source, a loss computed on the central labeled site's data, with respect to local parameters. Clients only receive the scalar loss value, not the data or a computational graph, so that gradient is undefined. No mechanism is given to transmit source gradients or any other information that would make the update executable. As written, line 7 of Algorithm 1 cannot run in PyTorch or any standard framework. This is a load-bearing flaw: the accuracies in Tables 4 and 5 cannot be traced to the described procedure. There are secondary issues too. The source site (USM for ABIDE, GE for ADNI) is chosen after seeing baseline performance on the same data, which is selection bias. Hyperparameters lambda1 and lambda2 are grid-searched on the same test folds. No code or data are released, so the empirical claims are hard to verify independently. There is also a duplicated sentence in the introduction, which suggests a lack of careful proofreading. If the intended algorithm is something like 'the central site sends its gradient and clients add their local gradient,' that needs to be stated explicitly and verified. As presented, the paper is not reproducible. The combination of ideas still has value, and a referee who knows federated learning could help determine whether a viable protocol exists. But I would not cite this version, and I would not trust the experimental results until the update rule is clarified and the evaluation protocol is fixed. My recommendation: send it to peer review only to let experts assess whether the algorithm can be salvaged; do not accept it in its current form.","headline":"Interesting combination of federated learning and domain adaptation for multi-site fMRI, but the core training step as written cannot be executed by local clients, so the reported gains are not reproducible.","tokens_in":679,"tokens_out":937,"would_cite":false,"duration_ms":40182,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A federated learning framework for fMRI, DAFed, splits each subject's brain connectivity into shared and site-specific features and, by training adversarially across hospitals without sharing patient data, claims the best average…","keywords":["federated learning","domain adversarial training","functional connectivity","feature disentanglement","contrastive learning","resting-state fMRI","autism spectrum disorder","mild cognitive impairment"],"falsifier":"A direct way to settle the central claim is to run Algorithm 1 under its stated privacy constraints and log whether the source-loss gradient used in line 7 can be computed at a local site without access to the central data. If local sites only receive the broadcast parameters and loss value, then the loss gradient with respect to local parameters can be computed for the target term but not for the source term; a working implementation must therefore transmit something beyond parameters, and checking what is transmitted (loss scalar, gradient surrogate, or data) determines whether the privacy claim holds.","tokens_in":19653,"feed_emoji":"🧠","tokens_out":7393,"duration_ms":72052,"temperature":0.7,"pith_summary":"This paper claims that a federated learning framework can train a single classifier across multiple hospitals' fMRI datasets without moving patient data, and that this approach outperforms both single-site and earlier federated methods on autism spectrum disorder (ASD) and mild cognitive impairment (MCI) classification. On the ABIDE multi-site dataset, the labeled variant reaches an average accuracy of 0.764 and the unlabeled variant reaches 0.753, versus 0.709 for the strongest federated baseline; on ADNI-3 it reaches 0.718. The building blocks are feature disentanglement, which splits each subject's connectivity features into components shared across sites and components unique to each site, adversarial training that transfers knowledge from a labeled central site to unlabeled local sites, and contrastive learning that aligns the shared features. The paper also argues that its interpretability module recovers known brain regions and functional connections associated with ASD and MCI despite the noise added for privacy.","feed_headline":"Federated brain-network model beats single-site classifiers on autism and MCI","feed_subtitle":"Splitting brain features into shared and site-specific parts lifts multi-site accuracy to 76.4 percent on ABIDE and 71.8 percent on ADNI.","key_machinery":"The machinery has five pieces. A graph convolutional feature generator with Jumping Knowledge builds a spatiotemporal embedding from dynamic functional connectivity matrices. A feature disentangler splits that embedding into the domain-invariant component (features that stay consistent across sites) and the domain-specific component, with a Mutual Information Neural Estimator minimizing their overlap. Two discriminators then act adversarially: a classifier trained on the labeled source domain, and a domain identifier that tries to tell whether features came from the source or a target site, so the feature extractor must fool the identifier on the invariant component. A multi-head self-attention mechanism fuses the two disentangled components before classification, and a contrastive loss pulls each local site's invariant features toward the global model's previous invariant features while pushing away older local versions. Federated averaging with Gaussian noise on local updates completes each round, and the improved Score-CAM replaces gradient-based explanations with score-weighted channel masks so that the added noise does not corrupt the attribution maps.","core_discovery":"DAFed's central discovery is that the non-IID problem in multi-site fMRI can be handled by explicitly modeling each dataset's features as a mixture of a domain-invariant part (the same across scanners and institutions) and a domain-specific part (local peculiarities), and then training with an adversarial objective so the classifier relies on the invariant part while the local part is preserved. The authors report that this design lets one labeled site (the source) teach unlabeled sites (the targets) under federated constraints, and that on the ABIDE and ADNI-3 cohorts the method achieves the best average classification accuracy among the compared methods in both labeled and unlabeled variants. They further claim the approach is generalizable: the same framework works for ASD diagnosis on ABIDE and MCI detection on ADNI-3, and the improved Score-CAM identifies both common and site-specific brain regions that align with earlier findings on these disorders.","pith_inferences":["The paper does not specify how a local site obtains the gradient of the central site's loss that Algorithm 1 uses in its update step; a working deployment would likely need to broadcast the source loss value or a gradient surrogate, which would change the privacy profile.","Because the method always designates one labeled site as the source and treats others as targets, it is effectively a semi-supervised domain adaptation scheme wearing federated clothing, so its accuracy gains may track the quality of the adversarial alignment more than the federated averaging itself.","The reported brain regions (e.g., amygdala, inferior frontal gyrus, default-mode network nodes) are testable biomarkers: using the same pipeline with a different source site or a different atlas would show whether the identified regions are stable or an artifact of the chosen source domain."],"forward_implications":["Multi-site studies can pool fMRI knowledge without transmitting raw scans, since only noise-added model parameters move between sites.","Sites that have no labels can still obtain a working classifier, because the adversarial transfer trains the shared extractor on the labeled source and applies it to unlabeled targets.","Scanner and institution differences that break naive federated averaging can be absorbed by the disentanglement module, which is why the method reports higher and more stable accuracy than FedAvg, FedMoE, FedAlign, and FedCL.","The noise-tolerant Score-CAM module gives the collaborative model an interpretability mechanism, so the brain regions and connections it highlights can be inspected even with privacy noise present."],"supporting_citations":[{"why":"Supplies the multi-site fMRI federated learning and domain adaptation setup that DAFed extends and compares against.","marker":"[6]"},{"why":"Provides the FedAvg baseline and the communication-efficient federated averaging mechanism that DAFed builds on.","marker":"[16]"},{"why":"Provides the FedCL federated contrastive learning baseline, which is the strongest federated comparator in the experiments.","marker":"[9]"},{"why":"Supplies the ABIDE multi-site autism dataset used for the ASD classification experiments.","marker":"[20]"},{"why":"Supplies the ADNI-3 multi-scanner dataset used for the MCI classification experiments.","marker":"[21]"},{"why":"Provides the Mutual Information Neural Estimator used to compute the mutual-information loss that enforces feature disentanglement.","marker":"[23]"},{"why":"Provides the domain-adversarial training objective that DAFed adapts to align labeled source and unlabeled target sites.","marker":"[32]"},{"why":"Supplies the Jumping Knowledge network design used to concatenate graph convolution layer outputs in the feature generator.","marker":"[22]"},{"why":"Provides the multi-head self-attention mechanism used to integrate domain-invariant and domain-specific features before classification.","marker":"[24]"},{"why":"Supplies the Score-CAM method that the improved, gradient-free interpretability module builds on.","marker":"[26]"}],"fun_headline_variants":["Disentangling brain features in federated learning boosts autism and MCI accuracy","DAFed: adversarial federated learning tackles non-IID fMRI across sites","Privacy-preserving federated brain network outperforms single-site classifiers","Splitting domain-invariant and site-specific features sharpens multi-site fMRI","Federated learning with disentangled brain features lifts ASD and MCI diagnosis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loop assumes that each local site can compute the gradient of the central site's loss with respect to its own model parameters, even though the central site's data and the source loss never leave the central site, and the paper gives no mechanism for obtaining that gradient.","fun_headline_variants_meta":{"raw":{"variants":["Disentangling brain features in federated learning boosts autism and MCI accuracy","DAFed: adversarial federated learning tackles non-IID fMRI across sites","Privacy-preserving federated brain network outperforms single-site classifiers","Splitting domain-invariant and site-specific features sharpens multi-site fMRI","Federated learning with disentangled brain features lifts ASD and MCI diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1567,"prompt_tokens":983,"completion_tokens":584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":599,"tokens_out":584,"duration_ms":7138,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:07:23.028233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct way to settle the central claim is to run Algorithm 1 under its stated privacy constraints and log whether the source-loss gradient used in line 7 can be computed at a local site without access to the central data. If local sites only receive the broadcast parameters and loss value, then the loss gradient with respect to local parameters can be computed for the target term but not for the source term; a working implementation must therefore transmit something beyond parameters, and checking what is transmitted (loss scalar, gradient surrogate, or data) determines whether the privacy claim holds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multi-site fMRI federated learning and domain adaptation setup that DAFed extends and compares against."},{"cited_title":"Communication -efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Provides the FedAvg baseline and the communication-efficient federated averaging mechanism that DAFed builds on."},{"cited_title":"FedCL: Federated contrastive learning for multi- center medical image classification","cited_arxiv_id":null,"evidence_quote":"Provides the FedCL federated contrastive learning baseline, which is the strongest federated comparator in the experiments."},{"cited_title":"The autism brain imaging data exchange: towards a large -scale evaluation of the intrinsic brain architecture in autism","cited_arxiv_id":null,"evidence_quote":"Supplies the ABIDE multi-site autism dataset used for the ASD classification experiments."},{"cited_title":"The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods","cited_arxiv_id":null,"evidence_quote":"Supplies the ADNI-3 multi-scanner dataset used for the MCI classification experiments."},{"cited_title":"Mutual Information Neural Estimation","cited_arxiv_id":null,"evidence_quote":"Provides the Mutual Information Neural Estimator used to compute the mutual-information loss that enforces feature disentanglement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the domain-adversarial training objective that DAFed adapts to align labeled source and unlabeled target sites."},{"cited_title":"Representation learning on graphs with jumping knowledge networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Jumping Knowledge network design used to concatenate graph convolution layer outputs in the feature generator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multi-head self-attention mechanism used to integrate domain-invariant and domain-specific features before classification."},{"cited_title":"Score -CAM: Score -weighted visual explanations for convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Score-CAM method that the improved, gradient-free interpretability module builds on."}],"review_version":1}