{"id":"1da86693-1b18-4611-89b2-8fb6abf59113","arxiv_id":"2607.29531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A multi-view, multi-source graph domain-adaptation framework with hyperbolic residual encoding reports 73.60% mean accuracy and 71.90% AUC across seven unlabeled rs-fMRI target sites for MDD identification.","lead":"This paper combines three types of brain-connectivity graphs with a multi-source domain-adaptation network to classify depression from resting-state fMRI across seven hospital sites, reporting 73.60% mean accuracy and 71.90% AUC. A generalist should read it as a test of whether unlabeled data from a new site can be used transductively to adapt a depression classifier without relabeling.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars or repeated-seed statistics; with target-site n as small as 30, the claimed 5.93-point ACC advantage and 'highest ACC on all seven sites' may be within chance variation.","rationale":"The reader's weakest assumption (pseudo-label confirmation bias) is a legitimate mechanism-level concern, but I find the statistical reliability of the headline numbers to be the most load-bearing issue. The paper's core claim is quantitative: 73.60% ACC, 71.90% AUC, and superiority over H2MSDA on every target site. Without repeated runs, the small target-site sample sizes (Site6: 30, Site12: 38) make these claims vulnerable to chance. A 2-sample difference on Site6 decides the 'highest ACC on all seven sites' claim. The absence of error bars is not just a reporting gap; it directly undermines the confidence in the central result. The pseudo-label loop, by contrast, is a potential failure mode that the empirical results could already refute if the method performs well; but we cannot know if it performs well consistently without variance estimates. I agree with the reader's overall CONDITIONAL verdict, but I would prioritize adding repeated-seed statistics and confidence intervals over addressing the pseudo-label mechanism. The paper does include a code link and honest acknowledgment of the transductive limitation, which are positive signs; the missing statistics are the concrete, testable gap.","tokens_in":10468,"tokens_out":6803,"duration_ms":74748,"concrete_test":"Run the complete seven-site protocol with 10 random seeds (different initializations and data orders), compute per-site and macro ACC/AUC mean ± std, and a paired bootstrap 95% CI over sites for the Ours minus H2MSDA difference. Also report accuracy on Site6 and Site12 per seed. If the CI includes 0 or any per-site advantage reverses in sign across seeds, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the method achieves 73.60% mean ACC / 71.90% AUC and beats H2MSDA by 5.93 ACC points across all seven target sites. For this to land, the reported numbers must be stable and the per-site advantages must not be noise. The paper reports a single run per configuration (no seeds, no standard deviations, no significance tests), while the smallest target sites have n=30 (Site6) and n=38 (Site12). At Site6, the 6.67-point ACC gap over H2MSDA corresponds to just 2 of 30 samples. There is also no description of how hyperparameters (λ_ss, λ_st, λ_adv, λ_im, λ_pl, τ schedule, etc.) were selected; if any were tuned on the target sites, the numbers are further inflated. This is more load-bearing than the pseudo-label confirmation-bias concern: even if the pseudo-label loop is unbiased, the headline superiority is not established without variability estimates. The paper's own conclusion restricts evidence to the transductive protocol, so the 'effective generalization' wording in the abstract is not supported by independent-site validation; but the numbers themselves are the load-bearing part.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-source, multi-view unsupervised graph domain adaptation framework for cross-site MDD identification from rs-fMRI. Three functional-connectivity views (Pearson correlation, sparse representation, Granger causality) are encoded by separate graph attention networks; pairwise cross-view interactions are integrated through a dual-stream adaptive fusion, followed by a Poincaré-ball hyperbolic residual encoder. Adaptation is driven by class-wise Cauchy–Schwarz alignment, adversarial learning, information maximization, and confidence-aware pseudo-labeling in a two-stage protocol. On seven target sites from REST-meta-MDD, the paper reports 73.60% mean accuracy and 71.90% mean AUC, with the highest per-site ACC on all seven sites and a 5.93-point mean ACC improvement over the H2MSDA baseline. The authors also include source-setting sensitivity, single/dual-source analysis, and ablations, and they release source code.","tokens_in":10851,"tokens_out":2325,"duration_ms":29648,"significance":"If the reported results are stable, the paper would make a useful contribution by jointly addressing multi-source and multi-view shifts in a transductive cross-site brain-imaging setting. The architectural components — especially the explicit modeling of pairwise cross-view interactions and the post-fusion hyperbolic residual — are well motivated, and the code release is a concrete strength for reproducibility. However, the central empirical claim is currently supported by single-run numbers without variance estimates or significance tests, and several target sites have very small sample sizes. The evidence is also explicitly transductive, so the abstract's wording about 'effective generalization under heterogeneous acquisition conditions' overstates what is demonstrated. The contribution is potentially solid but needs additional experimental rigor before publication.","major_comments":[{"comment":"The headline claim rests entirely on Table 2, but the paper reports only one run per method/site, with no standard deviations, confidence intervals, or significance tests. Target sites are small (Site6 n=30, Site12 n=38), so per-site differences are fragile: the 6.67-point ACC advantage at Site6 corresponds to roughly 2 of 30 samples, and the claimed 'highest ACC on all seven sites' may be within chance variation. Please provide results over multiple random seeds (at least 5–10), report mean±std per site and macro-averaged, and include paired tests versus H2MSDA and AUF-A. This is load-bearing because the superiority claim depends on the per-site gaps being non-noise.","section":"Table 2 and Section 5.2"},{"comment":"Several free hyperparameters are not reported: the loss weights λ_ss, λ_st, λ_adv, λ_im, λ_pl and their schedules, the pseudo-label confidence threshold τ and its schedule, the RBF kernel bandwidth in Eq. (17), the Poincaré-ball curvature c, and the graph sparsity criterion. The text fixes λ_SR=0.01 and GCM lag order=1 but not the others. Please report all hyperparameter values and, critically, the selection protocol. If any hyperparameter was tuned using target-site labels or by inspecting target-domain performance, the transductive unsupervised claim would be violated; if it was selected on a held-out validation split, describe the split. Without this, the reported numbers are not independently reproducible.","section":"Equations (2), (23) and Section 5.2"},{"comment":"The pseudo-label loop is a potential source of confirmation bias: class-wise source–target alignment and the pseudo-label loss both use the model's own confident target predictions, and the paper provides no analysis of pseudo-label quality or a correction mechanism. Early biased predictions can be reinforced by the alignment and pseudo-supervision, inflating apparent adaptation quality. Please report pseudo-label accuracy and coverage across adaptation iterations (or equivalent diagnostics), and discuss whether the confidence threshold schedule adequately mitigates this. If it is not mitigated, an independent-site (non-transductive) evaluation would be needed to separate the loop's actual benefit from self-reinforcement.","section":"Equations (18) and (22)"},{"comment":"The abstract states that the framework 'demonstrat[es] effective generalization under heterogeneous acquisition conditions,' but the conclusion explicitly limits the evidence to the transductive protocol: all target samples are used for adaptation, and no independent-site validation is provided. This is not an internal inconsistency, but the wording oversells the result. Please revise the abstract and Section 5.1 to say 'transductive adaptation performance' rather than generalization, or add an independent-site evaluation (e.g., train on Site20/21, adapt on a subset of one target site, and evaluate on held-out subjects from that site). The distinction matters for the clinical claim.","section":"Abstract and Section 6"}],"minor_comments":[{"comment":"The Pearson correlation view uses the absolute value of the correlation, but the text does not state whether sign information is intentionally discarded. Please clarify the rationale and whether the sign was tested in ablations.","section":"Equation (1)"},{"comment":"The Granger causality construction is described only in general terms. Specify whether bivariate or multivariate Granger causality was used, how the lag was chosen, and how the residual-variance reduction was computed. Also state whether the GCM matrix was symmetrized.","section":"Section 3.2"},{"comment":"In the ablation table, removing the hyperbolic encoder reduces AUC by 17.34 points (0.7190 vs 0.5456), which is a much larger effect than the ACC drop. A brief explanation of this asymmetry would help the reader interpret the hyperbolic module's role.","section":"Table 5"},{"comment":"The t-SNE visualization is presented as qualitative evidence, which is appropriate. Please state the t-SNE parameters (perplexity, number of iterations) and, if possible, provide a reproducibility note for the exact visualization.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The technical design is coherent and the code release is a plus, but the empirical core needs repeated-seed statistics and full hyperparameter reporting before the superiority claims can be evaluated. The pseudo-label confirmation-bias concern is real but secondary to the missing variance estimates. I would not reject: the issues are fixable within the manuscript's scope by additional experiments and reporting. I also recommend that the editor ask the authors to reconcile the abstract's 'generalization' language with the transductive protocol, which is stated accurately only in the conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: the paper is a coherent engineering contribution that bundles three FC views, pairwise cross-view fusion, hyperbolic residual encoding, and multi-source UDA into one architecture for cross-site MDD classification, but its central empirical claim—beating H2MSDA by 5.93 ACC points—is not yet established because there are no repeated seeds, error bars, or significance tests, and some target sites have only 30–38 samples.\n\nWhat's actually new: I haven't seen this exact combination in the literature. The framing of joint multi-view plus multi-source adaptation is legitimate, and the dual-stream fusion with pairwise interactions plus post-fusion hyperbolic encoding is a sensible design. The paper uses the public REST-meta-MDD dataset, reports per-site numbers, and ships a code link. The ablations are informative: removing DA or the hyperbolic encoder drops ACC substantially, and the single-source configurations show sensitivity to source selection. Those are useful empirical observations. The paper also explicitly restricts its evidence to the transductive protocol in the conclusion, which is honest.\n\nSoft spots, in order of importance. First, the absence of variability estimates is load-bearing. A macro-averaged ACC advantage of 5.93 points over H2MSDA is meaningful only if it is stable. With n=30 at Site6, a 6.67-point gap is 2 out of 30 subjects; no repeated seeds or standard deviations means the per-site 'highest ACC' claims could be within chance variation. Second, the hyperparameters are underreported: loss weights, schedules, confidence threshold τ, RBF bandwidth, curvature c, sparsity criterion. If any of these were tuned on the target sites, the numbers are inflated; the paper doesn't say. Third, the pseudo-label loop is a genuine concern: class-wise Cauchy–Schwarz alignment (Eq. 18) and pseudo-label loss (Eq. 22) use the model's own confident target predictions, so early bias can reinforce itself. The paper provides no correction or bound. That's a real limitation, but I'd rate it secondary to the missing variability estimates, because even a bias-free pseudo-label loop leaves the headline superiority unsubstantiated. Fourth, the abstract's 'effective generalization under heterogeneous acquisition conditions' overstates a transductive protocol; the conclusion hedges properly, the abstract doesn't.\n\nWho this is for: researchers working on transfer learning for rs-fMRI or multi-view graph UDA. It's a credible architectural proposal, but it needs a substantial revision before the accuracy claims are usable.\n\nRecommendation: send it to peer review, but the reviewers should demand repeated-seed statistics with error bars, full hyperparameter disclosure, and either an independent held-out-site evaluation or softened language. If the code is actually runnable, that helps; the paper deserves referee time, not a desk reject.","headline":"A coherent new architecture for multi-view multi-source MDD adaptation whose central accuracy claim needs error bars and full hyperparameter disclosure before it can be believed.","tokens_in":11305,"tokens_out":2212,"would_cite":false,"duration_ms":24318,"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":"Jointly modeling three functional-connectivity views with two labeled source sites and adapting to unlabeled target sites improves cross-site MDD identification, reaching 73.60% mean accuracy and 71.90% AUC over seven sites.","keywords":["domain adaptation","multi-view learning","functional connectivity","rs-fMRI","major depressive disorder","graph attention networks","hyperbolic embedding","Cauchy–Schwarz divergence"],"falsifier":"Hold out one entire acquisition site: run the same two-stage adaptation with the other seven sites as sources/targets and test on the held-out site with no adaptation on it. If accuracy on the held-out site falls to the level of the no-adaptation baseline (around 65.6% ACC), the reported gains come from transductive use of target data, not from transferable alignment.","tokens_in":10411,"feed_emoji":"🧠","tokens_out":5618,"duration_ms":57998,"temperature":0.7,"pith_summary":"The paper tries to establish that cross-site identification of major depressive disorder from resting-state fMRI can be improved by treating the problem as a multi-source, multi-view, unsupervised domain-adaptation task instead of handling multi-view connectome learning and cross-site adaptation separately. It builds three complementary functional-connectivity graphs—Pearson correlation, sparse representation, and Granger causality—encodes each with its own graph attention network, fuses them through pairwise cross-view interactions and a sample-adaptive dual stream, and refines the fused representation in a Poincaré-ball hyperbolic residual encoder. Class-wise Cauchy–Schwarz alignment then reduces discrepancies between the two labeled source sites and between each source and confident pseudo-labeled target samples, alongside adversarial, information-maximization, and pseudo-label objectives. Across seven independently adapted unlabeled target sites, the framework reports a macro-averaged 73.60% accuracy and 71.90% AUC, surpassing the strongest evaluated baseline by 5.93 and 4.75 percentage points. A sympathetic reader would take this as evidence that heterogeneous connectivity views and source identities should be preserved and aligned jointly rather than pooled or adapted independently.","feed_headline":"Depression model adapts to 7 new brain-scan sites at 73.6% accuracy","feed_subtitle":"Jointly aligning three fMRI connectivity views and two source sites beats single-source baselines by nearly six points.","key_machinery":"The load-bearing machinery is class-wise Cauchy–Schwarz alignment. The Cauchy–Schwarz divergence, computed with an RBF kernel, measures distribution discrepancy between feature sets; applied per class across the two source sites (Eq. 16) and between each source and confident target pseudo-label groups (Eq. 18), it is the mechanism that preserves class structure while pulling domains together. This is carried by the two-stage optimization, where labeled sources are pretrained and then the unlabeled target participates in adaptation; the confidence-aware pseudo-label loss provides the class-wise target groupings. The post-fusion hyperbolic residual encoder (Poincaré-ball constrained, with expo","core_discovery":"On its own terms, the paper's central claim is that a unified architecture can jointly handle inter-site distribution shifts, heterogeneous functional-connectivity views, and source-to-source differences. The authors construct three view-specific graph attention encoders, form pairwise cross-view interaction features, and combine them with the original view embeddings through a sample-adaptive dual-stream fusion. A post-fusion hyperbolic residual encoder, constrained to the Poincaré ball and read out through tangent-space maps, refines the fused subject-level representation. The adaptation objective couples class-wise Cauchy–Schwarz divergence between the two sources and between each source","pith_inferences":["My inference: the reported evaluation is transductive—each target site's unlabeled data participates in adaptation—so the numbers are not a direct measure of how the model would behave on a brand-new site never seen during adaptation; an independent-site holdout test (train/adapt on seven sites, test on an eighth) would be a stricter check.","My inference: the confidence-aware pseudo-label loop could suffer from confirmation bias if early pseudo-labels inherit source-site biases; a useful stress test would be to corrupt a fraction of confident pseudo-labels or vary the threshold schedule and observe whether accuracy degrades smoothly or abruptly.","My inference: because target sites vary widely in size and class balance, class-wise Cauchy–Schwarz estimates for small classes may be noisy; an extension would be to weight class terms by sample size or use robust kernel estimators.","My inference: the hyperbolic residual encoder's benefit could be disentangled from extra nonlinear capacity by swapping it for a Euclidean residual gate of matched parameter count; the paper does not report this comparison."],"forward_implications":["If the framework generalizes as reported, a new acquisition site's unlabeled scans could be used to adapt an MDD classifier without manual labeling at that site.","Preserving the two source sites as distinct sources outperformed pooling them (73.60% vs. 70.12% for the stronger single source), implying that source identity should be kept rather than collapsed.","Removing the domain-adaptation objective drops mean ACC from 73.60% to 65.61% and AUC from 71.90% to 52.58%, indicating the alignment terms, not the architecture alone, drive transfer.","Removing the hyperbolic encoder drops ACC to 64.86%, and the best single-view variant reaches only 70.34%, so multi-view fusion and curvature-aware refinement each contribute."],"fun_headline_variants":["Brain-scan AI adapts across 7 sites with 73.6% accuracy","MDD model fuses 3 fMRI views, adapts to 7 sites","Hyperbolic encoding lifts cross-site MDD accuracy to 73.6%","Multi-view adaptation: 73.6% MDD accuracy across 7 sites","Depression classifier jumps 6 points with joint view-source alignment"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework assumes that confident target predictions used as pseudo-labels are sufficiently accurate to guide class-wise alignment; if early pseudo-labels are biased by source-site distribution, the alignment and pseudo-supervision loop can reinforce that bias rather than correct it.","fun_headline_variants_meta":{"raw":{"variants":["Brain-scan AI adapts across 7 sites with 73.6% accuracy","MDD model fuses 3 fMRI views, adapts to 7 sites","Hyperbolic encoding lifts cross-site MDD accuracy to 73.6%","Multi-view adaptation: 73.6% MDD accuracy across 7 sites","Depression classifier jumps 6 points with joint view-source alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1535,"prompt_tokens":775,"completion_tokens":760,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":658}},"tokens_in":519,"tokens_out":760,"duration_ms":7778,"temperature":1.0,"reasoning_tokens":658,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:08:12.946848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out one entire acquisition site: run the same two-stage adaptation with the other seven sites as sources/targets and test on the held-out site with no adaptation on it. If accuracy on the held-out site falls to the level of the no-adaptation baseline (around 65.6% ACC), the reported gains come from transductive use of target data, not from transferable alignment.","supporting_citations":[],"review_version":1}