{"id":"f974f832-79d5-4ced-a465-a77fce127b7a","arxiv_id":"2602.10528","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Swap-adversarial learning with inter-subject channel swapping is claimed to improve cross-subject, cross-device, and cross-dataset PD classification, but the supporting evidence is limited by small samples, per-setting tuning, and a time-confounded benchmark.","lead":"A new framework combining cross-subject channel swapping with domain-adversarial training is claimed to improve Parkinson's disease classification across ECoG subjects, recording devices, and public EEG datasets. The paper also introduces a rat ECoG benchmark, but the class labels are time-locked to before/after stimulation, raising a serious confound.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MOCOP labels are time-locked to pre- vs post-stimulation with no sham control; models may learn temporal drift, so the central generalization claim is unsupported.","rationale":"The reader's strongest concern identifies the same load-bearing flaw: the MOCOP class labels are perfectly time-locked to pre- and post-stimulation with no sham control, so the benchmark's scientific validity and all ECoG generalization results depend on an assumption that is not tested. This is not a matter of conflicting with consensus; it is an internal design gap. The paper's internal ablations and t-SNE evidence show ISBCS reduces subject-specific separability, which is plausible, but none of that addresses whether the label itself is a nuisance variable. The claim 'consistently outperforms all baselines across all settings' is also weakened by per-setting grid search over lambda values and only six rats, but the time confound is more fundamental. A concrete test using sham animals or a within-class temporal split would settle whether time drift alone can explain the results. Until then, the paper's central claims are not adequately supported, and the reader's REJECT verdict stands.","tokens_in":21900,"tokens_out":3450,"duration_ms":41617,"concrete_test":"Obtain the raw ECoG recordings from the original Shin et al. study (which MOCOP is derived from) and run the exact SAF pipeline using the same pre-processing and hyperparameters on (a) any sham or unstimulated control animals recorded at the same pre/post timepoints, or if no such animals exist, (b) a within-class temporal split of the unstimulated recordings (e.g., first vs. last half of the pre-stimulation period). If the classifier achieves significantly above-chance accuracy in either setting, the labels are separable by time alone, confirming that the benchmark's class labels are confounded with temporal drift rather than stimulation-specific PD state. If no control data exist, that absence itself means the confound cannot be ruled out.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The dataset's class labels are defined purely by recording time: Section 4.1.2 states that class 0 is the state without electrical stimulation and class 1 is the state after one week of continuous stimulation. No sham, unstimulated, or counterbalanced control is described. Because all class-0 recordings precede all class-1 recordings within each animal, the label is perfectly confounded with time. Any slow drift in electrode impedance, recording equipment, animal behavior, or environmental conditions over the week could produce the same spectral changes the paper attributes to stimulation-induced neural state. The reported PSD shifts (delta increase in wireless, broadband decrease in wired) are exactly the kind of global changes that temporal drift can generate. If the model learns this time-correlated nuisance, then the high accuracies in Experiments 1 and 2 (e.g., 98.9% wireless cross-subject) and the claimed cross-dataset generalization do not reflect PD-related neural state. This invalidates the 'first reproducible benchmark' contribution and the central claim that SAF learns task-relevant shared features. The absence of a control condition is a load-bearing gap, not a minor caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Swap-Adversarial Framework (SAF) for domain generalization in ECoG/EEG classification, combining preprocessing (band-pass, notch, ASR), an Inter-Subject Balanced Channel Swap (ISBCS) augmentation that swaps corresponding channels between subjects of the same class, and domain-adversarial learning with a gradient reversal layer and an entropy-based mutual-information penalty. The authors introduce MOCOP, an ECoG benchmark built from 6-OHDA-lesioned rat recordings, and report cross-subject, cross-modality (wireless/wired), ablation, and cross-dataset EEG experiments claiming consistent improvements over EEGNet and DMMR baselines, with the largest gains in the most variable environments. The central claims are that SAF learns task-relevant, subject-invariant features and that MOCOP is the first reproducible benchmark for ECoG-based PD prediction.","tokens_in":22089,"tokens_out":7343,"duration_ms":79382,"significance":"If the benchmark labels were valid and the reported gains were robust, the framework would be a useful contribution: it is a simple, modular recipe (preprocessing + data augmentation + adversarial training) that could transfer across brain-signal modalities, and MOCOP would fill a real gap as a public ECoG benchmark for PD prediction. The paper has concrete strengths: the ablation study supports complementarity of ISBCS and DAL, the method is evaluated under several shift scenarios, and the authors commit to releasing data and code. However, the label construction in MOCOP, the per-target hyperparameter selection, and the circular F-statistic evidence all undermine the central generalization and benchmark claims. As presented, the evidence does not establish that the model learns PD-related neural state rather than time-correlated nuisance variation.","major_comments":[{"comment":"The dataset labels are perfectly confounded with recording time. Class 0 is defined as the state without electrical stimulation and class 1 as the state after one week of continuous stimulation, with no sham control, unstimulated control group, or counterbalanced design described. Because all class-0 recordings precede all class-1 recordings within each animal, any slow drift in electrode impedance, recording equipment, behavior, or environment can produce the global spectral changes reported in Figure 7. The high accuracies in Experiments 1 and 2 (e.g., 98.9% wireless cross-subject) may therefore reflect classification of temporal drift rather than PD-related neural state. This is a load-bearing validity issue for MOCOP as a 'reproducible benchmark' and for the central generalization claim. The manuscript must show control analyses (e.g., shuffled-time labels, unstimulated animals) or s","section":"Section 4.1.2"},{"comment":"Hyperparameters lambda_MI and lambda_GRL are grid-searched and reported separately for each target rat and each transfer direction (e.g., Wireless Rat1: 0.001/0.001; Wireless Rat3: 3.334/1.25). If the validation split used to select these values includes any target-domain data, the 'unseen domain' evaluation is transductive and the comparison is optimistic. The paper must state exactly how validation folds were constructed relative to the source/target split. In addition, the EEG cross-dataset results in Figure 17 are reported as point estimates with no error bars or significance tests, so the claim that the method 'consistently outperforms all baselines across all settings' is not statistically supported.","section":"Section 5.3 / Table 3"},{"comment":"The F-statistic drop from 24.48 to 0.99 after ISBCS is a direct consequence of the augmentation construction: swapping channels between subjects mixes subject identities in the swapped channels by design, so between-subject variance in the feature space decreases mechanically. This is not independent evidence that task-relevant, subject-invariant features are learned. The passage should be reframed as a sanity check of the augmentation's intended effect, not as 'statistical evidence' for the method's generalization benefit. The circularity weakens the central claim that ISBCS improves generalization.","section":"Section 5.6"},{"comment":"Each ECoG setting contains only three source subjects; the cross-subject evaluation is at most a 2-vs-1 split repeated three times. The reported +/- values (e.g., 0.989 +/- 0.010 for wireless accuracy) are standard deviations across three fold results and are not a stable estimate of model performance. No significance tests are provided. With n=3, one easy or hard target subject can drive the average, so the claimed 41% improvement over EEGNet in wireless accuracy should be interpreted with this limitation explicitly acknowledged.","section":"Section 5.5, Experiments 1-2"},{"comment":"The 'first reproducible benchmark' claim is not currently verifiable. MOCOP is a re-annotation and re-partitioning of recordings from Shin et al. (2025), and the manuscript provides no data DOI, accession number, or code link ('will be made publicly available upon publication'). Reproducibility cannot be assessed. Together with the label confound identified in Section 4.1.2, the benchmark contribution as stated is not supportable.","section":"Contribution 1 / Section 4.1"}],"minor_comments":[{"comment":"The abstract's first sentence swaps EEG and ECoG definitions: 'Electroencephalography (ECoG)' and 'electrocorticography (EEG)'. The title also says 'Electroencephalography-Based' while the paper is about ECoG. Please correct throughout.","section":"Abstract and Title"},{"comment":"L_MI is defined as the entropy of the domain classifier's predicted subject distribution H(h_psi(z)). This is not the mutual information between z and s; it is a proxy. Please rename it as a conditional-entropy proxy or derive the actual mutual information.","section":"Eq. (3)"},{"comment":"The baseline name is inconsistent: Section 5.1 lists 'DDMR' while Figures 13-17 and the Results sections use 'DMMR'. Choose one spelling.","section":"Section 5.1 and Figures"},{"comment":"Section 4.2.1 says ASR was not applied to the EEG datasets, but Section 5.3 says clean_asr() was used for both ECoG and EEG data. Clarify which preprocessing the EEG models actually received.","section":"Section 4.2.1 vs Section 5.3"},{"comment":"ISBCS assumes channel indices correspond functionally across subjects, described as 'guided by a brain map,' but no anatomical registration or channel-mapping procedure is described. Please make the correspondence explicit or discuss its validity.","section":"Section 3.3.1"},{"comment":"The silhouette score histograms and t-SNE plots are used to support claims of inter-subject separability, but no confidence intervals or permutation tests are provided. Please temper the wording or add quantitative uncertainty.","section":"Sections 4.1.3 and 4.2.2"}],"recommendation":"reject","confidential_remarks":"The label confound alone is sufficient to reject the manuscript in its current form: the central benchmark and generalization claims are unsupported. A resubmission could be considered if the authors obtain or analyze control data (e.g., sham/unstimulated recordings or time-shuffled labels) and substantially strengthen the statistical evaluation, but that is beyond the scope of a standard revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a swap-adversarial framework (ISBCS plus domain-adversarial learning) for cross-subject and cross-dataset generalization in ECoG/EEG, and introduces MOCOP, a re-annotated rat ECoG benchmark built from the authors' prior recordings. The method is simple and plausible: same-class channel swapping between subjects followed by domain-adversarial training is a reasonable extension of mixup-style augmentation, and the internal ablations are consistent with both components helping. The EEG cross-dataset results (UI to UNM and back) are also interesting and are not subject to the same confound as the ECoG data.\n\nThe soft spot is the benchmark design. Class labels are defined by recording time — pre-stimulation vs. after one week of stimulation — with no sham, counterbalanced, or unstimulated control. Any slow drift in electrodes, behavior, or environment over that week is perfectly confounded with the label. The reported PSD changes (delta increase in wireless, broadband decrease in wired) are exactly the kind of global shifts temporal drift produces. If the model is learning drift rather than PD-related neural state, the 'first reproducible benchmark' contribution and the cross-subject generalization claims lose their force. This is load-bearing, not a minor caveat.\n\nOther issues are secondary but real: only three source subjects per ECoG modality, per-setting grid-selected hyperparameters (Table 3) that can absorb variance, no error bars on the EEG cross-dataset results, and an abstract that promises cross-session experiments that never appear in the text. The F-statistic drop after ISBCS is a direct consequence of the augmentation, so it isn't independent evidence. And the public release is only promised 'upon publication,' so 'reproducible benchmark' is currently unverifiable.\n\nAll that said, the paper is clearly written, the method is coherent, and the idea of combining channel-swapping augmentation with adversarial domain suppression is worth discussion. The EEG experiments suggest the framework has some general value beyond the ECoG benchmark. My recommendation: this deserves a serious referee, because the dataset and method could be useful and the confound needs to be exposed. But I would not accept it as is — the authors need to address the temporal-drift problem (e.g., a time-shuffled label control, or behavioral/electrophysiological evidence that the changes are specifically stimulation-driven) and fix the overclaims in the abstract before the central claims are credible.","headline":"Plausible method and a potentially useful ECoG benchmark, but the benchmark's time-locked labels with no sham control undercut the central generalization claim.","tokens_in":22659,"tokens_out":2206,"would_cite":false,"duration_ms":25537,"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":"A channel-swap plus adversarial training recipe lets ECoG models generalize to unseen subjects in Parkinson's prediction, outperforming baselines across every setting tested.","keywords":["ECoG","EEG","Parkinson's disease","domain generalization","adversarial learning","data augmentation","channel swapping","benchmark dataset"],"falsifier":"Train the framework on MOCOP with class labels shuffled within each subject's time sequence, or collect a sham-stimulated cohort where the same electrode implant and handling occur without active stimulation; if accuracy remains near 0.99, the model is reading time-linked artifacts rather than PD-related neural state.","tokens_in":21719,"feed_emoji":"🧠","tokens_out":2708,"duration_ms":29132,"temperature":0.7,"pith_summary":"The paper claims that a two-step recipe — swapping functionally matched channels between same-class subjects, then adversarially training a subject discriminator — makes deep models generalize to unseen subjects in ECoG-based Parkinson's disease classification. The authors argue that the swap weakens subject-specific structural cues before adversarial training, so the adversarial step removes remaining subject bias without destroying task-relevant signal. They report consistent gains over EEGNet and DMMR baselines across cross-subject, cross-modality (wireless-wired), and cross-dataset (EEG benchmark) settings, with the largest gains in highly variable environments. They also introduce MOCOP, an open ECoG benchmark from six 6-OHDA rats, intended as the first reproducible benchmark for ECoG-based PD prediction.","feed_headline":"Swap-adversarial training beats baselines on unseen ECoG subjects","feed_subtitle":"Channel swap plus adversarial de-biasing lifts Parkinson's classification across rats, modalities, and EEG datasets.","key_machinery":"Inter-Subject Balanced Channel Swap (ISBCS) is the data-level engine: with probability p per channel, it swaps the signal of a channel between two same-class samples from different subjects, guided by a brain map to pair functionally corresponding channels. This is coupled with domain-adversarial learning through a Gradient Reversal Layer (GRL) and a domain classifier predicting subject identity from features, plus a mutual-information penalty on the predicted subject distribution. The backbone is EEGNet, a compact CNN for brain signals.","core_discovery":"On its own terms, the paper establishes that inter-subject channel swapping (ISBCS), applied at the data level before domain-adversarial learning, reduces the F-statistic of between-subject vs within-subject feature variance from 24.48 to 0.99, and that the full framework outperforms baselines in every setting tested. The central claim is that the two components are complementary: ISBCS weakens subject-specific structure that adversarial learning can then remove without harming task-relevant features, yielding subject-invariant representations that generalize to unseen subjects and unseen recording environments.","pith_inferences":["The claim that corresponding channels preserve task-relevant signal relies on an anatomical correspondence that the paper does not verify; a testable extension is to swap random channels and compare accuracy to see how much the brain-map guidance matters.","The benchmark's labels are time-confounded (pre-stimulation versus after one week of stimulation), so part of the reported accuracy may reflect slow electrode drift or behavioral state rather than PD-specific neural change; a sham-stimulated control group would separate these.","The F-statistic reduction shows ISBCS makes subjects indistinguishable in feature space, but it does not show that the remaining features carry the class information; a class-conditional or mutual-information measure would sharpen the claim.","Because the framework generalizes from ECoG to EEG, it may also generalize to other electrophysiological modalities such as local field potentials or single-unit recordings, which could be tested with existing public datasets."],"forward_implications":["If correct, cross-subject ECoG classification becomes practical with small cohorts: channel swapping requires no new data and the gains are largest exactly where subject variability is high.","The framework transfers across modalities — training on wireless recordings and testing on wired (or vice versa) works, and it also transfers across EEG datasets from different institutions.","The MOCOP dataset, once public, would give the ECoG community a standard benchmark with reproducible subject splits and a trained baseline, enabling fair comparisons that the paper argues were previously missing.","The reduction in F-statistic from 24.48 to 0.99 provides a measurable target for other augmentation methods: subject separability in feature space can be driven to near zero without necessarily destroying task information."],"fun_headline_variants":["Channel swap plus adversarial learning generalize Parkinson's ECoG","Swap-adversarial framework tames subject variability in Parkinson's ECoG","Channel swap and adversarial learning boost cross-dataset Parkinson's ECoG","New swap-adversarial method generalizes ECoG across subjects and datasets","Adversarial channel swap improves Parkinson's ECoG generalization"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The two classes are defined by recording time — before versus one week after stimulation — with no sham or counterbalanced control, so any slow drift in electrodes, behavior, or environment is perfectly confounded with the label; if the model exploits that drift, the central generalization claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Channel swap plus adversarial learning generalize Parkinson's ECoG","Swap-adversarial framework tames subject variability in Parkinson's ECoG","Channel swap and adversarial learning boost cross-dataset Parkinson's ECoG","New swap-adversarial method generalizes ECoG across subjects and datasets","Adversarial channel swap improves Parkinson's ECoG generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000798,"raw_usage":{"total_tokens":3356,"prompt_tokens":763,"completion_tokens":2593,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2500}},"tokens_in":507,"tokens_out":2593,"duration_ms":20168,"temperature":1.0,"reasoning_tokens":2500,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:04:11.344793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the framework on MOCOP with class labels shuffled within each subject's time sequence, or collect a sham-stimulated cohort where the same electrode implant and handling occur without active stimulation; if accuracy remains near 0.99, the model is reading time-linked artifacts rather than PD-related neural state.","supporting_citations":[],"review_version":1}