{"id":"bad66ae5-5fd1-4fb7-b019-6882ded1c60f","arxiv_id":"2507.14147","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A GCNN built from EEG spectral coherence and electrode distance classifies insomnia with 70% window accuracy, and channel-ablation tests highlight C4-P4, F4-C4 and C4-A1 as the most informative channels.","lead":"This paper trains a graph convolutional neural network on EEG brain networks to tell people with insomnia from healthy sleepers, reaching about 70% accuracy per window and 68% per person. It also reports that removing certain electrodes, C4-P4, F4-C4 and C4-A1, hurts performance most, pointing to brain regions known to be altered in insomnia.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (1)'s random-coherence model is invalid as written (C_random > 1 at D=0) and the constant k is never specified, so all edge weights and the reported channel-importance ranking are not reproducible and may not represent brain connectivity.","rationale":"The reader's weakest assumption identifies exactly the point on which the central argument rests: the graph edges, and therefore the brain-network interpretation and channel-importance ranking, are built from Eq. (1). I considered replacing this with the evaluation-uncertainty concern (16 subjects, no confidence intervals, data-selected 50-s window), which is real, but it is secondary: even a perfectly executed cross-validation would not make the brain-modeling claim sound if the connectivity matrix is computed from an internally inconsistent formula. The exp(1 - Dij/k) form at D=0 produces a value greater than 1, which is not a valid coherence, and the constant k is absent, making the procedure unreproducible. This is not a disagreement with external consensus; it is an internal correctness issue that can be settled by recomputation. The reader's conditional verdict already demands full reproducibility artifacts and confidence intervals; the concrete test above would additionally settle whether the brain-network results survive a corrected random-coherence model. No verdict change is needed beyond the published conditional decision.","tokens_in":9129,"tokens_out":9433,"duration_ms":114885,"concrete_test":"Re-run the 50-s window experiments (Tables 1 and 3) after replacing Eq. (1) with the standard reduced-coherence form C_random = exp(-Dij/k), using the value of k given in Nunez et al. [19] and also with k fit on a small held-out validation split. If the window-level accuracy moves by more than a few percentage points, or if the channel-importance ordering (C4-P4, F4-C4, C4-A1) changes, the reported results are not robust to the random-coherence specification. The authors should also state the units of Dij and the value or fitting rule for k.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the random-coherence correction in Eq. (1) and its use in Eq. (3). As written, C_random^ij = exp(1 - Dij/k) exceeds 1 whenever Dij < k (and equals e approximately 2.72 at Dij = 0), which is impossible for a coherence measure that should lie in [0, 1]. The constant k is never given a numerical value or fitting procedure, so every connectivity matrix, every edge weight, and the classification accuracy and channel-importance ranking in Tables 1-3 are not reproducible and could change substantially under a correct or differently calibrated correction. The same applies to the unweighted addition of Dij in Eq. (3), which mixes a physical distance with a dimensionless coherence without a stated scale. The paper's stated conclusion that the three most important channels are positioned near brain regions known to exhibit atypical functional connectivity depends on edge weights that may be artifacts of this mis-specified formula. The empirical accuracy could in principle survive the correction, but the brain-modelling claim and channel ranking cannot be evaluated until Eq. (1) is corrected and k reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a graph convolutional neural network model to classify insomnia from continuous sleep EEG. Brain graphs are built from five channels, using spectral coherence corrected by a random-coherence model and augmented with geodesic distance as edge weights, and power spectral density in six frequency bands as node features. The authors report 5-fold subject-independent cross-validation over 16 subjects (9 insomnia, 7 control), finding that a 50-second window yields 70.1% window-level and 67.7% subject-level accuracy (Table 1), and that combining coherence with distance improves accuracy (Table 2). Channel-ablation experiments identify C4-P4, F4-C4 and C4-A1 as the most important channels (Table 3).","tokens_in":9322,"tokens_out":5339,"duration_ms":58634,"significance":"If the central claims hold, this is an early demonstration of GCNN-based brain modelling for insomnia from long-duration EEG, and the subject-independent cross-validation is a genuine strength. The systematic window-length comparison and channel-ablation protocol are useful, and the use of continuous ~13-hour recordings goes beyond typical short sleep-onset windows in prior work. However, the validity of the connectivity model is not established, and the small sample and lack of statistical inference substantially weaken the quantitative conclusions. The paper currently provides a proof-of-concept rather than a reliable clinical or neurobiological result.","major_comments":[{"comment":"The random-coherence model in Eq. (1) is invalid as written: C_random^ij = exp(1 − D_ij/k) exceeds 1 whenever D_ij < k, reaching e at D_ij = 0, which is impossible for a coherence quantity. The constant k in [19] is never specified numerically or fitted, so the reduced coherence in Eq. (2) and every edge weight in Eq. (3) are not reproducible and may take invalid negative values. Since Tables 1–3 depend entirely on these edge weights, the classification and channel-importance results cannot currently be interpreted as brain-connectivity measures. Please correct the formula (or use the actual random-coherence estimator from Nunez et al.), report k and the normalization procedure, and show that the resulting edge weights lie in [0,1].","section":"§3.3, Eq. (1)–(2)"},{"comment":"Eq. (3) adds a physical geodesic distance D_ij, measured in centimetres or similar, to a dimensionless coherence value C_ij without any stated scaling or units. The subsequent normalisation to [0,1] is not described, so it is unclear whether the improvement attributed to the spatial term in Table 2 reflects a genuine connectivity enhancement or just a rescaling of the graph. Provide the normalisation formula and a sensitivity analysis over the relative scaling of coherence and distance.","section":"§3.3, Eq. (3)"},{"comment":"The 50-second window is selected because it gives the highest accuracy on the same subject-independent 5-fold cross-validation folds that are then used to report the final performance. This makes the reported accuracy optimistic and does not account for model selection. Use nested cross-validation or pre-specify the window length, and report the variance across folds and iterations.","section":"§4.1, Table 1"},{"comment":"The study includes only 16 subjects (9 insomnia, 7 control), and no error bars, confidence intervals, or significance tests are reported anywhere. In this setting, the small differences between conditions in Tables 1–3 (e.g., 70.1% vs. 68.5% for Fp2-F4 omission) may easily be within chance. Report per-subject accuracies, exact binomial or permutation tests, and compare against at least one classical baseline (e.g., SVM or logistic regression on the same features) to establish that the GCNN architecture contributes beyond the features.","section":"§3.1, §4"},{"comment":"In the channel-importance experiment, omitting a channel reduces the graph from five to four nodes and changes the model architecture, so the resulting drop in accuracy conflates information loss with a change in model capacity/input size. To support the neurobiological interpretation, the ablation should be compared against removing a randomly chosen channel in a permutation test, or the model should be re-run with a four-node graph using different subsets to separate the effect of losing a particular channel.","section":"§4.2, Table 3"}],"minor_comments":[{"comment":"The abstract reports subject-level accuracy as 68%, while Table 1 and the main text report 67.7%; please make these consistent.","section":"Abstract vs. Table 1"},{"comment":"The GCNN hyperparameters (number of layers, hidden dimensions, dropout, batch size, number of epochs, early stopping) are not reported; include them for reproducibility.","section":"§3.4 and §4"},{"comment":"The exact coherence estimator from [5] and the frequency range or band over which coherence is computed are not specified; the connectivity matrix could differ substantially across bands.","section":"§3.3"},{"comment":"Clarify whether artefact rejection or sleep-stage selection was performed on the ~13-hour recordings, since the whole recording is used without further description.","section":"§3.2"},{"comment":"Figures 3–5 show no error bars or spread; given the small sample, add per-fold or per-subject variability.","section":"Figures 3–5"},{"comment":"Table 3 evaluates only the five channels in the montage; the conclusion about the 'three most important channels' should be explicitly limited to this five-channel setup.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper against your notes and the stress-test. The stress-test is right: Eq. (1) is invalid as written. C_random^ij = exp(1 - Dij/k) is greater than 1 whenever Dij < k, and at Dij = 0 it gives e ≈ 2.72, which is impossible for a coherence measure. The constant k is never specified, so every edge weight in Eq. (3) — and the channel-importance ranking the paper leans on — is built on an artifact. The paper cannot be evaluated as a brain-connectivity study until that formula is corrected and k is reported. That is the main thing you need to know.\n\nWhat the paper does have: a clean subject-independent 5-fold cross-validation, a sensible window-length sweep, and a channel-ablation protocol. It is the first GCNN application to insomnia on long-duration CAP EEG, and the 70% window-level / 68% subject-level numbers, if they survived a corrected connectivity measure, would be a plausible exploratory result. The writing is clear and the CAP database is public, so the experiment is replicable in principle.\n\nBut beyond Eq. (1): only 16 subjects (9 insomnia, 7 control) with no confidence intervals or significance tests; the 50-second window is selected on the same folds used to report accuracy; there is no baseline classifier; and hyperparameters and code are not released. The channel-ablation experiment tests only five channels, so calling three of them 'most important' from a ranking of five is a stretch. The anatomical interpretation in Section 4.2 is plausible but rests on the invalid edge weights.\n\nMy verdict: this is an interesting exploratory application, but the load-bearing connectivity equation is wrong as written and the evidence base is too thin to support the conclusions. I would not send it to peer review in its current form. It needs a corrected formula, a specified k, error bars, and at least one baseline. If the authors fix the formula and the numbers move, the revised version might be worth a look.","headline":"Small-sample insomnia EEG-GCNN paper with a broken coherence formula; the channel-ranking result is not trustable as brain connectivity.","tokens_in":9859,"tokens_out":2356,"would_cite":false,"duration_ms":25403,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph convolutional network built from spectral-coherence brain graphs and power-spectral-density node features can classify insomnia from continuous sleep EEG, reaching 70.1% window-level and 66.7% subject-level accuracy, and…","keywords":["insomnia detection","electroencephalography (EEG)","graph convolutional neural network","functional connectivity","spectral coherence","power spectral density","channel ablation","sleep brain networks"],"falsifier":"Record scalp EEG from a known dipole source, or run a realistic head-model simulation with known true connectivity, and check whether the edge weights from this formula recover that ground truth; a miscalibrated $k$ would leave residual distance-dependent coherence. Alternatively, rerun the classification on the same data after replacing the random-coherence correction with a Laplacian spatial filter, and look for a large change in accuracy or in the C4-P4 ranking.","tokens_in":8899,"feed_emoji":"🧠","tokens_out":10804,"duration_ms":104454,"temperature":0.7,"pith_summary":"The paper argues that insomnia leaves a learnable signature in long-duration sleep EEG. If each 50-second segment is turned into a small brain graph whose edges are spectral coherence corrected for volume conduction and augmented with electrode distance, and whose nodes carry band power-spectral density, a graph convolutional network can separate insomnia patients from healthy sleepers at 70.1% window-level and 66.7% subject-level accuracy. In the best configuration, precision, recall, and F1 all sit near 0.70, which matters because existing brain models cover Alzheimer's disease and epilepsy but not insomnia, and because a model that points to specific EEG channels could support non-drug diagnosis and treatment assessment. The paper positions the result as a proof-of-concept that functional-connectivity graphs plus spatial information, rather than raw waveforms, carry enough signal for objective insomnia screening.","feed_headline":"Sleep-EEG brain graphs spot insomnia at 70 percent","feed_subtitle":"Network models on coherence and distance trace insomnia's EEG signature and flag central electrodes","key_machinery":"The load-bearing object is the weighted brain graph built per EEG window. Nodes are EEG channels, acting as proxies for cortical regions; edges are spectral coherence values reduced by a random-coherence correction and then linearly combined with geodesic electrode distance, with the result normalized to $[0,1]$. The random-coherence correction, $C^{(random)}_{ij}=\\exp(1 - D_{ij}/k)$, is the mechanism intended to remove volume conduction, and the additive distance term is what distinguishes this connectivity estimate from plain coherence. Node features are per-channel power spectral density in six bands computed by Welch's method, and a three-layer graph convolutional network with global add pooling and dense layers maps these graphs to class probabilities.","core_discovery":"On its own terms, the paper's central discovery is that a domain-guided graph convolutional neural network, using edge weights $C'_{ij}=C^{(computed)}_{ij}-\\exp(1 - D_{ij}/k)+D_{ij}$ normalized to $[0,1]$ and node features from Welch power-spectral-density estimates in six frequency bands, classifies insomnia from continuous sleep EEG at 70.1% window-level and 66.7% subject-level accuracy with a 50-second non-overlapping window. With spectral coherence alone, the same pipeline reaches about 64% window-level and 57% subject-level accuracy, so the additive distance term improves the connectivity estimate. Ablating individual channels shows that removing C4-P4, F4-C4, or C4-A1 degrades accuracy the most, and the paper connects those electrodes to brain regions with previously reported atypical functional connectivity in insomnia. The claim is that a graph representation of functional connectivity, not the raw multichannel time series, is sufficient to carry insomnia-related information in long sleep recordings.","pith_inferences":["Editorial inference: the channel-importance ranking is a model-internal attribution, not direct evidence of brain-region causality, unless the connectivity estimate is validated against source-localized data.","A testable extension would redo the classification with a Laplacian spatial filter instead of the Eq. (1) random-coherence correction; if accuracy or the C4-P4 ranking changes sharply, the reported result depends on that correction rather than on the brain signal.","Another testable extension is sensitivity analysis over the constant $k$ in Eq. (1); if the best window-level accuracy moves by more than a few points when $k$ is re-estimated on this dataset, the edge-weight scale is not robust.","An inference for future work: feeding the same graphs to a simpler classifier, such as an SVM on graph metrics, would show how much of the 70% comes from the graph features themselves versus the graph convolutional network."],"forward_implications":["A 50-second window is the best temporal scale in this EEG setting; shorter windows weaken the signal and longer windows likely straddle sleep micro-structure cycle boundaries, reducing accuracy.","Combining spectral coherence with electrode distance is strictly better than coherence alone, indicating that spatial geometry carries complementary discriminative information for insomnia.","The channel-importance ranking gives a concrete compression target: a simplified device should keep C4-P4, F4-C4, and C4-A1 before other channels.","Subject-level accuracy remains above chance but below window-level accuracy, so the learned features are not explained by a single subject, although patient-level variability still matters.","The same graph-construction and classification pipeline should transfer to other sleep pathologies if the connectivity signature is similarly learnable."],"supporting_citations":[{"why":"Supplies the random-coherence correction formula $C^{(random)}_{ij}=\\exp(1-D_{ij}/k)$ used to subtract volume conduction from measured coherence.","marker":"[19]"},{"why":"Defines the coherence function used to compute the measured coherence values $C^{(computed)}_{ij}$.","marker":"[5]"},{"why":"Contributes the idea of augmenting EEG functional connectivity with spatial electrode distance for graph-based diagnosis, which Eq. (3) adopts.","marker":"[24]"},{"why":"Provides the graph convolutional operator that the model architecture is built on.","marker":"[15]"},{"why":"Defines the sleep micro-structure cycle whose typical sub-minute duration motivates the window-length comparison.","marker":"[22]"},{"why":"Supplies clinical evidence of insomnia-related functional connectivity in sensory-motor regions, used to interpret channel importance.","marker":"[14]"},{"why":"Shows increased insula coactivation with salience networks in insomnia, supporting the importance of central electrodes.","marker":"[8]"},{"why":"Reports increased insular connectivity in primary insomnia, another clinical anchor for the channel-ablation results.","marker":"[25]"},{"why":"Earlier sleep-EEG classification of psycho-physiological insomnia with mutual-information brain networks, used as evaluation context and baseline.","marker":"[3]"}],"fun_headline_variants":["Sleep EEG graphs spot insomnia with 70% accuracy","Graph neural nets read insomnia in sleep EEG","Insomnia EEG decoded via brain connectivity graphs","Brain-graph model flags insomnia from sleep EEG","GCNN on sleep EEG traces insomnia to key electrodes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on assuming that the random-coherence correction in Eq. (1), with its fixed constant $k$ taken from earlier work, adequately removes volume conduction from the EEG coherence; if that correction is miscalibrated, the brain graphs are artifacts and the accuracy and channel ranking are not measuring brain connectivity.","fun_headline_variants_meta":{"raw":{"variants":["Sleep EEG graphs spot insomnia with 70% accuracy","Graph neural nets read insomnia in sleep EEG","Insomnia EEG decoded via brain connectivity graphs","Brain-graph model flags insomnia from sleep EEG","GCNN on sleep EEG traces insomnia to key electrodes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1301,"prompt_tokens":1007,"completion_tokens":294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":223}},"tokens_in":623,"tokens_out":294,"duration_ms":4402,"temperature":1.0,"reasoning_tokens":223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:43:56.020020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record scalp EEG from a known dipole source, or run a realistic head-model simulation with known true connectivity, and check whether the edge weights from this formula recover that ground truth; a miscalibrated $k$ would leave residual distance-dependent coherence. Alternatively, rerun the classification on the same data after replacing the random-coherence correction with a Laplacian spatial filter, and look for a large change in accuracy or in the C4-P4 ranking.","supporting_citations":[{"cited_title":"Electroencephalography and clinical neurophysiology103(5), 499–515 (1997)","cited_arxiv_id":null,"evidence_quote":"Supplies the random-coherence correction formula $C^{(random)}_{ij}=\\exp(1-D_{ij}/k)$ used to subtract volume conduction from measured coherence."},{"cited_title":"John Wiley & Sons (2011)","cited_arxiv_id":null,"evidence_quote":"Defines the coherence function used to compute the measured coherence values $C^{(computed)}_{ij}$."},{"cited_title":"In: Machine Learning for Health","cited_arxiv_id":null,"evidence_quote":"Contributes the idea of augmenting EEG functional connectivity with spatial electrode distance for graph-based diagnosis, which Eq. (3) adopts."},{"cited_title":"Sleep medicine2(6), 537–554 (2001)","cited_arxiv_id":null,"evidence_quote":"Defines the sleep micro-structure cycle whose typical sub-minute duration motivates the window-length comparison."},{"cited_title":"Neuroreport 24(5), 233–240 (2013)","cited_arxiv_id":null,"evidence_quote":"Supplies clinical evidence of insomnia-related functional connectivity in sensory-motor regions, used to interpret channel importance."},{"cited_title":"Biological Psychology97, 1–8 (2014)","cited_arxiv_id":null,"evidence_quote":"Shows increased insula coactivation with salience networks in insomnia, supporting the importance of central electrodes."},{"cited_title":"European Radiology27, 3703–3709 (2017)","cited_arxiv_id":null,"evidence_quote":"Reports increased insular connectivity in primary insomnia, another clinical anchor for the channel-ablation results."},{"cited_title":"Journal of medical systems39, 1–10 (2015)","cited_arxiv_id":null,"evidence_quote":"Earlier sleep-EEG classification of psycho-physiological insomnia with mutual-information brain networks, used as evaluation context and baseline."}],"review_version":1}