{"id":"6a666f9e-11f9-4d1e-9aa6-ff0ef303b332","arxiv_id":"2509.10524","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FENet combines time- and frequency-domain brain network views in a CCA-based self-supervised framework, improving psychiatric disorder classification accuracy on ABIDE and ADHD-200 over prior graph SSL methods.","lead":"This paper introduces FENet, a self-supervised deep learning method that reads fMRI brain scans in both time and frequency domains to detect psychiatric disorders like autism and ADHD from small datasets. Across two public datasets, it beats existing graph-based methods, especially when only 10-20% of labels are available.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random site-mixed 5-fold CV cannot support SOTA or high-frequency biomarker claims; a site-hold-out with training-only pretraining is needed.","rationale":"The reader's weakest assumption -- that site effects can be ignored under random splits -- is the most load-bearing concern because it directly determines whether the reported performance and the high-frequency biological conclusion are valid. I agree with the conditional verdict: the proposed architecture is plausible, the complexity analysis is useful, and the core ideas are not internally contradictory, but the evaluation as reported does not rule out the dominant confound. My additional observation about pretraining on the entire dataset (Section 5.4 and Algorithm 1) reinforces the same concern: if test subjects are included in unsupervised pretraining, the benchmark is transductive and not directly comparable to standard inductive SSL evaluation. The proposed concrete test -- site-hold-out with training-only pretraining -- would settle both issues at once. I do not see a need to move beyond the reader's conditional judgment; the paper should be accepted only after such verification or with claims softened to the specific random-split protocol.","tokens_in":21705,"tokens_out":7641,"duration_ms":82511,"concrete_test":"Run leave-one-site-out (or site-stratified) cross-validation on ABIDE and ADHD-200 with the same FENet pipeline, and additionally re-run the reported 5-fold protocol with SSL pretraining restricted to each fold's training subjects only. Compare FENet against the three strongest baselines (GATE, A-GCL, CCA-SSG) using paired significance tests. If FENet's accuracy/F1 advantage shrinks to within noise or reverses on held-out sites, the performance and high-frequency claims are artifacts of site mixing and/or transductive pretraining.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FENet is the strongest SSL method on ABIDE and ADHD-200 rests on the assumption that the random stratified 5-fold split in Section 5.4 isolates disorder-related signal from acquisition site. ABIDE aggregates 17 sites and ADHD-200 aggregates 8 sites; scanner, protocol, and motion effects are strongly site-correlated. With random splits, the same sites appear in both training and test folds, so a model can score well by recognizing site-specific artifacts rather than psychiatric disorder. This threatens the headline SOTA improvements in Section 6.1 and the high-frequency biomarker interpretation in Section 6.3.1. The risk is compounded by the stated pretraining protocol: Section 5.4 says the SSL model is trained on 'the entire dataset' before fine-tuning, so if this includes validation/test subjects, the encoder has transductive access to test-distribution information; it is unclear whether baselines receive the same treatment. No site-stratified or site-hold-out experiments are reported, and no significance tests are provided; the reported 2-4% gains have overlapping standard deviations and are not yet shown to reflect disease rather than site or leakage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FENet, a self-supervised framework for fMRI-based detection of psychiatric disorders (ASD and ADHD). It constructs time- and frequency-domain brain graph views from ROI-level BOLD signals, encodes them with separate GNN encoders (TGNN and an FGNN using graph filters and Fourier graph operators), and aligns the two views via a soft-CCA objective with decorrelation terms (Eq. 11). Experiments on ABIDE and ADHD-200 are reported against supervised and SSL baselines, with claims of state-of-the-art accuracy/AUC/F1, data efficiency under 10-20% label rates, and an ablation suggesting that high-frequency components are the most informative for disorder detection.","tokens_in":22021,"tokens_out":5506,"duration_ms":49627,"significance":"If the empirical claims withstand scrutiny, FENet would be a useful contribution: it appears to be the first brain-network SSL method to explicitly combine time- and frequency-domain views under a non-contrastive CCA-style objective, and the FGO-based encoder provides a computationally attractive way to incorporate spectral information. The paper is clearly written, includes thorough ablations of the learning objective and hyperparameters, and provides code links for all baselines. However, the evaluation protocol currently has site-confounding and pretraining-separation issues that make the headline performance and high-frequency biomarker claims uncertain; these are addressable with additional experiments. The paper does not provide code for FENet itself, which limits reproducibility of the proposed method.","major_comments":[{"comment":"The random stratified 5-fold split is not site-aware. ABIDE aggregates 17 acquisition sites and ADHD-200 aggregates 8 sites, and scanner/protocol/motion effects are strongly site-correlated. With random splits the same sites appear in both training and test folds, so the model can score well by recognizing site-specific artifacts rather than psychiatric disorder. No site harmonization, site-stratified folds, or site-hold-out is reported. This threatens the SOTA claim in §6.1 and the frequency-band interpretation in §6.3.1. Please add a site-hold-out experiment (e.g., leave-site-out) with SSL pretraining confined to training sites, and report per-site or site-group metrics. At minimum, report site composition of folds and demonstrate that the model does not exploit site identity.","section":"§5.4 and §5.1"},{"comment":"The pretraining description states: 'the model is trained in a fully unsupervised manner for 200 iterations on the entire dataset.' If 'entire dataset' includes the test fold, the SSL pretraining stage has transductive access to the test distribution (though not the labels), which can inflate fine-tuning performance—especially in the low-label experiments of §6.2. It must be clarified whether pretraining is restricted to training folds and whether all SSL baselines follow exactly the same train/test separation during pretraining. If the current protocol is transductive, the data-efficiency claim in Figure 7 should be re-run with pretraining on training folds only.","section":"§5.4"},{"comment":"No statistical significance tests are reported. Several head-to-head differences are within one standard deviation of the repeated runs: e.g., ABIDE ACC 62.5±3.9 vs GATE 59.8±3.4; ADHD AUC 69.7±4.4 vs A-GCL 69.0±3.2; ADHD F1 67.0±5.2 vs A-GCL 64.7±3.9. Reporting only mean±std is insufficient to support 'FENet outperforms all baseline methods.' Please add paired tests across the same fold/seed structure (e.g., paired bootstrap, McNemar, or corrected resampled t-test) or report confidence intervals.","section":"§6.1, Table 2"},{"comment":"The conclusion that high-frequency components are critical is inferred from the ablation FENet_FH having the best accuracy, but the frequency thresholds λ_L and λ_H were 'optimized' on the same datasets and the final dual-filter architecture was selected after observing these results. This makes the high-frequency biomarker interpretation vulnerable to selection bias. Please pre-specify thresholds from the literature (e.g., [19,44]) or evaluate on an independent hold-out/nested CV, and report the dual-filter model against single-band variants under fixed thresholds.","section":"§6.3.1 and §5.4"}],"minor_comments":[{"comment":"The initialization line reads 'Initialize: The model parameters θF and θF'; the second should be θT.","section":"Algorithm 1, line 3"},{"comment":"The phrase 'recall is slightly lower than A-GCL' understates the ABIDE gap: 63.3±4.8 vs 71.6±12.9. Please quantify the gap and acknowledge the high variance of the A-GCL recall estimate.","section":"§6.1"},{"comment":"The data-efficiency curves show no error bars or variance information. Given the repeated-seed protocol described in §5.4, please include error bars or state explicitly that the plotted points are means over seeds.","section":"Figure 7"},{"comment":"The definition of S_{0:p} and the role of the complex-valued bias b_p is under-specified. Clarify how complex values are handled before being fed into the real-valued objective in Eq. (11).","section":"§4.3.2, Eq. (10)"},{"comment":"'200 iterations' is ambiguous; specify whether this means 200 epochs or 200 optimizer steps.","section":"§5.4"},{"comment":"Table 1 lists K_L and K_H as frequency threshold parameters, but the text uses λ_L and λ_H. Align the notation for consistency.","section":"Table 1 and §5.4"},{"comment":"No code release for FENet is mentioned, while links are provided for all baseline methods. Making the proposed method's implementation available would strengthen the reproducibility of the benchmark-style claims.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The main concern is not the method itself but the evaluation protocol. Site confounding is a well-known issue in ABIDE/ADHD-200 studies, and the paper currently does not address it; the site-hold-out or site-stratified analysis is essential before the SOTA claim can be accepted. The self-reported limitation in §7 about cross-database generalization should be extended to within-dataset site effects. I do not see a novelty or scope problem for ACM TCH; the topic fits the journal. The paper is otherwise well organized and the method is plausible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable SSL method for fMRI that combines graph-spectral frequency views with CCA-style cross-domain alignment, and it shows consistent gains over a strong baseline at low label rates. The evaluation, however, has a site-mixing problem and a potential pretraining leak, so the SOTA and high-frequency biomarker claims are not yet established.\n\nWhat's new: the specific architecture—time and frequency brain graphs, separate TGNN and FGNN encoders, and a CCA objective that aligns the two domains—is not in the cited literature. GATE and UCGL use temporal views only, CCA-SSG is generic graph SSL, and Zhang et al. is time-series. The FGO-layer frequency encoder is a sensible efficiency choice, and the ablations show that frequency features help, especially high-frequency filters. The data-efficiency experiments (10-20% labeled) show larger gains over CCA-SSG, which is encouraging.\n\nWhere it gets soft: ABIDE and ADHD-200 are multi-site datasets. Random stratified 5-fold CV means the same sites appear in train and test, so the model can exploit site-specific artifacts rather than disorder. No site-holdout or harmonization is reported, and the 2-4% improvements come with overlapping standard deviations and no significance tests. More troubling, Section 5.4 says pretraining is done on 'the entire dataset' before fine-tuning. If that includes validation/test subjects, the encoder has transductive access to test information; it's unclear whether baselines get the same treatment. That alone makes the current SOTA claim unpersuasive. The high-frequency conclusion is drawn from an ablation where the filter thresholds were chosen on the same data, so it's self-referential.\n\nThese are fixable. The method is not obviously wrong, and the idea deserves a proper test. But the paper needs site-stratified evaluation, a clean pretraining protocol, significance tests, and code release before the claims can be trusted.\n\nRecommendation: send to peer review; the referees should require these revisions. It's a worthwhile contribution if the results survive a stricter protocol.","headline":"Solid idea, shaky evaluation: the frequency-enhanced SSL method is a real combination, but site mixing and a possible pretraining leak undermine the SOTA and biomarker claims.","tokens_in":22471,"tokens_out":2260,"would_cite":false,"duration_ms":22144,"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":"FENet claims that a self-supervised model jointly learning time-domain and frequency-domain brain-graph views detects ASD and ADHD from fMRI more accurately than existing supervised and self-supervised baselines, especially when only a frac","keywords":["psychiatric disorder detection","fMRI","self-supervised learning","brain networks","frequency-domain analysis","graph neural networks","canonical correlation analysis","data efficiency"],"falsifier":"Run FENet with site-hold-out evaluation: train on a subset of acquisition sites and test on held-out sites from ABIDE and ADHD-200. If accuracy drops to near chance or the high-frequency components lose their predictive value, site effects rather than disorder-related frequency patterns carry the reported result. Alternatively, train a linear probe to predict the acquisition site from the learned FENet representation; high site-prediction accuracy would indicate the representation encodes site confounds.","tokens_in":21663,"feed_emoji":"🧠","tokens_out":4433,"duration_ms":45536,"temperature":0.7,"pith_summary":"The paper is trying to establish that fMRI-based psychiatric disorder detection can be made data-efficient by explicitly teaching a model to use both temporal and spectral views of the same brain scan. It proposes FENet, a self-supervised framework that builds two brain graphs from one fMRI time series—one in the time domain and one in the graph-frequency domain—and aligns their learned representations with a consistency objective. The authors claim FENet outperforms all compared graph-based supervised and self-supervised methods on the ABIDE and ADHD-200 datasets while using only 20% of labels for fine-tuning, and that high-frequency components carry a particularly strong disorder signal. If true, this would make fMRI-based screening more practical in clinical settings where labeled psychiatric data are scarce.","feed_headline":"Frequency-aware model tops ASD and ADHD fMRI benchmarks","feed_subtitle":"FENet fuses time and spectral brain graphs and beats baselines using only 20% of labels.","key_machinery":"The central mechanism is multi-view brain graph augmentation: the same ROI-level BOLD time series is represented once as a time-domain connectivity graph and once as a graph-frequency graph via the Graph Fourier Transform, sharing the same adjacency matrix. Two domain-specific encoders process these views—a time-domain GCN and a frequency-domain FGNN that uses a graph filter over Laplacian eigenvalues plus Fourier Graph Operator layers to keep computation log-linear. The learning objective is a CCA-style domain consistency loss: it minimizes the distance between time and frequency representations of the same subject while decorrelating each representation's dimensions, balancing alignment wi","core_discovery":"FENet integrates time-domain functional connectivity with frequency-domain spectral features by transforming the same BOLD signals through the Graph Fourier Transform, then encoding each view with a dedicated graph neural network. A domain-consistency-guided canonical correlation analysis objective pulls the two representations together while decorrelating each view's feature dimensions, producing a fused representation used for disorder classification. On ABIDE, FENet reports accuracy, AUC, and F1-score gains of 2.7%, 1.4%, and 4.9% over the strongest baseline; on ADHD-200 the gains are 3.4%, 0.7%, and 2.3%. Under 20% labeled data, it reports accuracy improvements of 6.5 and 11.2 percentage","pith_inferences":["A site-hold-out evaluation, training on some acquisition sites and testing on others, would clarify whether the reported gains reflect disorder-related signal or scanner/site differences; the authors themselves flag harmonization as future work.","The same time-frequency consistency objective could transfer to other spectral neuroimaging modalities such as EEG or MEG, but the paper only tests fMRI with the AAL atlas.","The ablation showing that mid-frequency components add little suggests a concrete design rule worth testing: drop the mid-band in the frequency encoder on new datasets and check whether accuracy holds.","Pretraining on large unlabelled fMRI repositories before fine-tuning could push the minimal-label boundary further; FENet currently pretrains on the same datasets it is evaluated on."],"forward_implications":["If FENet's claims hold, psychiatric-disorder classifiers can be trained with a fraction of the labels currently needed, reducing the cost of fMRI-based screening and diagnosis.","Frequency-domain views can be derived from the same scan without new data collection, so the approach can be added to existing brain-graph SSL pipelines.","The reported high-frequency dominance suggests that future diagnostic models should retain high-frequency graph components rather than treating them as noise.","The log-linear complexity of the frequency encoder means the method can scale to higher-resolution brain atlases with more ROIs.","The larger gains at 20% labeled data compared with CCA-SSG directly support the paper's data-efficiency claim rather than only its absolute accuracy."],"supporting_citations":[{"why":"Supplies the graph-frequency decomposition and the medical interpretation of low- versus high-frequency brain graph components that FENet relies on.","marker":"[15]"},{"why":"BrainGNN is the supervised ROI-aware GNN baseline that FENet must outperform and whose interpretability goal FENet extends.","marker":"[22]"},{"why":"BrainGB provides the standardized brain network construction and GNN benchmarking pipeline used as a baseline.","marker":"[8]"},{"why":"GATE is the closest prior CCA-based temporal SSL method for fMRI; FENet extends its CCA idea to cross-domain alignment.","marker":"[37]"},{"why":"CCA-SSG supplies the CCA-based self-supervised objective that FENet adapts with domain-specific trade-off coefficients.","marker":"[61]"},{"why":"FourierGNN contributes the Fourier Graph Operator and Green's kernel that make FENet's frequency encoder computationally efficient.","marker":"[58]"},{"why":"Introduces the time-frequency domain consistency concept that FENet's learning objective instantiates for brain graphs.","marker":"[64]"},{"why":"A-GCL is the strongest contrastive SSL baseline in the comparisons and the reference for the recall trade-off discussion.","marker":"[62]"},{"why":"Supplies the clinical evidence that high-frequency neural oscillations are modulated by task and linked to cognitive processing, supporting FENet's high-frequency emphasis.","marker":"[44]"}],"fun_headline_variants":["Frequency-enhanced brain graphs boost psychiatric fMRI detection","FENet: Fusing time and frequency brain signals for better disorder detection","Self-supervised frequency-enhanced brain networks edge out baselines","FENet beats ASD and ADHD benchmarks by fusing time and frequency brain graphs"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that random stratified five-fold splits of multi-site fMRI data make scanner and acquisition-site differences ignorable, so the measured classification performance reflects the disorder rather than the site.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-enhanced brain graphs boost psychiatric fMRI detection","FENet: Fusing time and frequency brain signals for better disorder detection","Self-supervised frequency-enhanced brain networks edge out baselines","FENet beats ASD and ADHD benchmarks by fusing time and frequency brain graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2789,"prompt_tokens":774,"completion_tokens":2015,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1954}},"tokens_in":518,"tokens_out":2015,"duration_ms":13985,"temperature":1.0,"reasoning_tokens":1954,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:33:59.614965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FENet with site-hold-out evaluation: train on a subset of acquisition sites and test on held-out sites from ABIDE and ADHD-200. If accuracy drops to near chance or the high-frequency components lose their predictive value, site effects rather than disorder-related frequency patterns carry the reported result. Alternatively, train a linear probe to predict the acquisition site from the learned FENet representation; high site-prediction accuracy would indicate the representation encodes site confounds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the graph-frequency decomposition and the medical interpretation of low- versus high-frequency brain graph components that FENet relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BrainGNN is the supervised ROI-aware GNN baseline that FENet must outperform and whose interpretability goal FENet extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BrainGB provides the standardized brain network construction and GNN benchmarking pipeline used as a baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GATE is the closest prior CCA-based temporal SSL method for fMRI; FENet extends its CCA idea to cross-domain alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CCA-SSG supplies the CCA-based self-supervised objective that FENet adapts with domain-specific trade-off coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FourierGNN contributes the Fourier Graph Operator and Green's kernel that make FENet's frequency encoder computationally efficient."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the time-frequency domain consistency concept that FENet's learning objective instantiates for brain graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A-GCL is the strongest contrastive SSL baseline in the comparisons and the reference for the recall trade-off discussion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the clinical evidence that high-frequency neural oscillations are modulated by task and linked to cognitive processing, supporting FENet's high-frequency emphasis."}],"review_version":1}