{"id":"f3ade385-9888-4712-a5d2-75b45939fb06","arxiv_id":"2509.07991","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A single deep model with regional zero-padding and dual local-global encoders classifies EEG from motor imagery, resting state, and driving fatigue, reaching 59.6% macro F1.","lead":"This paper proposes DLGE, a transformer-based model that classifies EEG from three different BCI tasks (motor imagery, resting state, driving fatigue) using one shared architecture and standardized brain-region channel padding. The authors report macro F1 around 59.6% on the combined eight-class problem, but the evaluation does not test the central claim of generalization to an unseen paradigm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation never holds out a paradigm and uses shuffled-segment CV, so the 'without retraining' claim is untested and likely inflated.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing flaw: the evaluation protocol cannot support the stated generalization claim. The paper's own text confirms the protocol: 'Finally, we shuffle all the samples' (Section II-B) and 'five-fold cross-validation' (Section II-D). This design conflates three distinct failures: (1) no paradigm is held out, so the model is never tested on an unfamiliar task type or channel layout; (2) segment-level shuffling violates independence because adjacent segments from the same continuous recording are temporally correlated, leaking subject/session identity into test folds; (3) the output layer is ficed to the eight classes of the three training paradigms, so the architecture is not even structurally capable of accommodating a new paradigm without modification. Any one of these would invalidate the 'without retraining' claim; together they make the empirical evidence irrelevant to the main contribution. Other weaknesses (no baselines, no code, no hyperparameter details) are secondary. I agree with the reader's REJECT verdict and recommend no change in the outcome.","tokens_in":13176,"tokens_out":5937,"duration_ms":73735,"concrete_test":"Run leave-one-paradigm-out (LOPO) cross-validation: train DLGE on samples from any two paradigms and test on the held-out third paradigm, ensuring that training folds use subject-level (not segment-level) splits to avoid temporal leakage. If macro F1 on the held-out paradigm does not significantly exceed the chance level appropriate for that paradigm's classes, the cross-paradigm generalization claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that DLGE generalizes across BCI paradigms without retraining, but the only evaluation (Section II-D) is 5-fold CV on all samples after 'we shuffle all the samples' (Section II-B). This test uses a pooled 8-class classifier with all three paradigms present in every training fold; no paradigm is ever held out, so the model is never required to transfer to an unseen paradigm or channel configuration. Adjacent non-overlapping 3-second segments from the same continuous recording (especially resting-state and driving-fatigue sessions) are highly correlated; random fold assignment places segments from the same subject/session into both training and test folds, leaking subject- and recording-specific information and inflating macro F1. The reported ~60% macro F1 (chance=12.5%) may reflect memorization of low-level recording artifacts rather than learnable cross-paradigm representations. Thus the abstract's claim of processing diverse BCI paradigms 'without retraining and retuning' is not supported by the reported evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DLGE, a dual local-global encoder for EEG classification across three BCI paradigms: motor imagery (4 classes), resting state (2 classes), and driving fatigue (2 classes), giving an 8-class pooled task. To handle heterogeneous channel configurations, the authors partition channels into 11 anatomical brain regions, standardize each region to 9 channels, and zero-pad missing channels. The local brain encoder applies temporal and spatial attention per region, the global brain encoder aggregates regional features, and a cross-task classifier outputs predictions. The authors report mean macro precision, recall, and F1 of 60.16%, 59.88%, and 59.56% from five-fold cross-validation after shuffling all samples from the three datasets, plus ablation, parameter sensitivity, and Grad-CAM analyses.","tokens_in":13464,"tokens_out":3837,"duration_ms":47393,"significance":"If the central claim were supported, the paper would make a useful contribution: a single model that handles multiple BCI paradigms with heterogeneous channel layouts, without per-paradigm retraining, would be relevant to practical brain-computer interfaces. The anatomical brain-region padding strategy and the local-global architecture are sensible design ideas, and the ablation study provides some insight into which components matter. However, the evaluation design does not test the headline claim. All three paradigms are present in every training fold, no paradigm is ever held out, and the random shuffle of continuous 3-second segments creates likely information leakage between training and test folds. The reported numbers are therefore not evidence of cross-paradigm generalization. The absence of baseline comparisons and statistical tests further limits the interpretability of the results.","major_comments":[{"comment":"The evaluation protocol does not test cross-paradigm generalization. The authors state 'Finally, we shuffle all the samples' (II-B) and then use five-fold CV (II-D) on the pooled eight-class data. Every fold contains training and test segments from all three paradigms, and no paradigm is ever held out. Consequently, the central claim in the Abstract ('processing diverse BCI paradigms without retraining and retuning') is untested: the model is only shown to classify a fixed set of eight paradigm-condition classes, with each paradigm represented in the training set. Moreover, because 3-second segments are cut from continuous recordings (5,184 MI; ~1,000 resting; 2,400 fatigue samples), adjacent non-overlapping segments from the same subject and session are highly correlated; random shuffling places correlated segments in both training and test folds, inflating all reported metrics. A valid","section":"Section II-B, II-D"},{"comment":"The reported macro F1 of ~59.56% is presented without any baseline or comparison. With eight classes chance is 12.5%, but no comparison is made to (i) a single-paradigm model retrained per paradigm, (ii) a simple spectral or Riemannian baseline, or (iii) a model trained on shuffled segments with subject-disjoint folds. Because the results in Table II are obtained under the leaky protocol of Major Comment 1, the numbers cannot be interpreted as evidence of generalizable cross-paradigm representations. The absence of confidence intervals or significance tests further weakens the claim.","section":"Section III-A, Table II"},{"comment":"In Eq. (6), the second projection is written q(r)=W_k v(r); it should be k(r)=W_k v(r). More substantively, the spatial interaction in Eq. (7) computes softmax over all ~C=9 channels per region, including zero-padded channels. Padded channels contribute zero-valued keys, which can dilute attention weights assigned to real channels. The authors mask padded positions in the GBE (Section II-C.2) but do not describe an equivalent mask in the LBE spatial attention. This should be clarified or corrected, since it affects the validity of the regional feature extraction for paradigms with fewer than nine channels in a region.","section":"Section II-C, Eqs. (6)-(9)"}],"minor_comments":[{"comment":"The text reports 'approximately 1 2%' for the standard deviation across folds; this should be '1-2%' or the exact value should be given. Table II also lacks standard deviations and confidence intervals.","section":"Section III-A"},{"comment":"The y-axis in the ablation figure starts at 47.5%, which visually exaggerates differences. Reporting effect sizes with error bars would be more informative.","section":"Fig. 3"},{"comment":"No code or data availability statement is provided, which hinders reproducibility. Given the novelty of the claim, releasing the preprocessing and evaluation scripts would be valuable.","section":"General"},{"comment":"The Grad-CAM visualizations are computed from a model that has access to paradigm-specific channel layouts and padding patterns; the authors should discuss whether these maps might reflect dataset artifacts rather than neurophysiological mechanisms.","section":"Section III-D"}],"recommendation":"reject","confidential_remarks":"The central evaluation flaw is fundamental: the paper claims cross-paradigm generalization without ever holding out a paradigm. Even a major revision would require re-running the entire experimental pipeline with leave-one-paradigm-out or subject-disjoint splits and adding baseline comparisons. As submitted, the evidence does not support the abstract's claim. The architectural ideas are worth pursuing, but the manuscript in its current form is not ready for journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one thing genuinely new: it tackles channel-configuration heterogeneity across BCI paradigms with an anatomical brain-region partition plus zero padding, and combines a local per-region encoder with a global encoder. That's a reasonable engineering contribution, and the ablation suggests the global encoder earns its keep—removing it drops macro F1 by ~7.7 points. The writing is clear, the method is described in reproducible detail, and the choice of three real paradigms (motor imagery, resting state, driving fatigue) is sensible.\n\nThe soft spot is the evaluation. The paper claims 'without retraining and retuning' for diverse paradigms, but no paradigm is ever held out. The 5-fold CV is run on samples shuffled across all three paradigms, so every training fold contains motor imagery, resting-state, and fatigue samples; the model never faces an unseen paradigm or an unseen channel layout. Worse, samples are 3-second non-overlapping segments from continuous recordings; random fold assignment leaks within-subject, within-session correlations into training, which can inflate macro F1 well above the 12.5% chance level. So the headline claim is untested. The reported 59.6% macro F1 may reflect low-level recording artifacts or subject identity rather than cross-paradigm representation.\n\nThere are also no baselines—no EEGNet, no Conformer, no within-paradigm comparison—so we can't tell whether the architecture adds anything over a generic transformer with padding. The parameter sensitivity and Grad-CAM analyses are internal consistency checks, not evidence of generalization.\n\nThat said, the problem is real and the padding strategy is a plausible building block. If the authors re-ran the evaluation with a leave-one-paradigm-out split, and ideally subject-wise folds within that, the paper would be worth taking seriously. As is, it's a promising proof-of-concept with a load-bearing evaluation flaw.\n\nRecommendation: I'd send it to peer review rather than desk-reject, because the idea and ablation are substantive and the flaw is fixable. But I'd tell the authors the evaluation must be redone before acceptance.","headline":"A reasonable cross-paradigm architecture, but the evaluation never holds out a paradigm, so the headline claim is untested.","tokens_in":13868,"tokens_out":1667,"would_cite":false,"duration_ms":18437,"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":"This paper claims that a dual local-global encoder, with anatomical brain-region padding, lets one EEG model classify three BCI paradigms without retraining.","keywords":["cross-BCI-paradigm classification","EEG decoding","brain-region partitioning","zero-padding channel standardization","dual local-global encoder","motor imagery","resting state","driving fatigue"],"falsifier":"Train DLGE on motor imagery and resting-state data only, then test it on all driving-fatigue data without any fatigue samples in the training folds. If the held-out-paradigm F1 falls to near chance, the central claim fails. As a second check, compare the current shuffled-segment result with a subject-disjoint split; a large drop would indicate that the model exploited within-recording segment similarity.","tokens_in":13097,"feed_emoji":"🧠","tokens_out":5107,"duration_ms":53832,"temperature":0.7,"pith_summary":"The paper tries to show that one deep-learning model can classify EEG from different BCI paradigms—motor imagery, resting state, and driving fatigue—even though these paradigms were recorded with different channel montages and involve different mental tasks. To bridge the channel mismatch, the model groups electrodes into 11 anatomical brain regions and zero-pads each region to a fixed 9-channel layout. A local brain encoder learns features shared across paradigms within each region, and a global brain encoder combines them into task-specific representations. On a pooled, shuffled test set the authors report macro precision 60.16%, recall 59.88%, and F1 59.56%. If this holds, it would mean a single trained model could be reused across BCI paradigms without retraining or re-tuning.","feed_headline":"One EEG model decodes three BCI paradigms at ~60% F1","feed_subtitle":"Brain-region padding plus dual local-global encoders lets one model skip per-paradigm retraining.","key_machinery":"The key object is the anatomical brain-region standardization plus the dual encoder stack. First, each sample's EEG channels are grouped into 11 fixed brain regions, and each region is padded to exactly 9 channels, with padded channels masked in later attention. The local brain encoder treats each channel's time-frequency matrix as a token sequence, applies a single-head transformer per channel as a channel embedder, computes spatial attention between channels inside a region, and condenses valid channels into one region embedding. The global brain encoder then applies multi-head self-attention over the 11 region embeddings, with padded regions masked by setting their attention scores to neg","core_discovery":"The central claim is that DLGE, a dual local-global encoder, achieves cross-BCI-paradigm classification without retraining. Channel heterogeneity is handled upstream of learning: electrodes from any montage are mapped to the same 11 brain regions, and missing positions are zero-padded, so the transformer-based feature extractor always sees a standardized regional input. Shared regional features are extracted by the local encoder using temporal attention per channel and spatial attention across channels; the global encoder then merges the 11 regional summaries into the high-level representation used by the classifier. Evaluated on the three paradigms together, the model reaches an average mac","pith_inferences":["The paper's random shuffled 5-fold cross-validation is a weaker test than holding out an entire paradigm; the strongest version of the claim, 'without retraining on a new paradigm,' would require training on two paradigms and testing on the third.","Because adjacent 3-second segments from the same continuous recording are correlated, the current split may overestimate cross-paradigm generalization; a subject-disjoint or recording-disjoint split would give a more conservative estimate.","The zero-padding strategy means different paradigms effectively have different numbers of valid channels per region; comparing performance against a fixed common subset of channels would clarify how much of the gain comes from padding versus genuine shared representations.","The regional contributions shown by the visualization are averaged over a classification model that sees all paradigms together; a per-paradigm analysis with a held-out paradigm would show whether the model learns paradigm-specific shortcuts or stable neural signatures."],"forward_implications":["A model trained once on several paradigms could accept EEG from a new device with a different channel montage, if its channels can be mapped into the 11 brain regions.","Cross-paradigm decoding would remove the need to store and maintain separate models for motor imagery, vigilance, and resting-state applications.","The small optimal depth (two attention layers) suggests the model can be kept light enough for portable BCI hardware.","The same regional-padding strategy could be applied to other EEG tasks with different channel densities, as long as an anatomical mapping is available.","The reported small fold-to-fold variation suggests stable training, which is a prerequisite for deploying a single universal decoder."],"supporting_citations":[{"why":"supplies the motor imagery paradigm data (four classes, 22 channels) used in the cross-paradigm evaluation.","marker":"[32]"},{"why":"supplies the resting-state paradigm data (eyes open vs closed, 64 channels).","marker":"[33]"},{"why":"supplies the driving fatigue paradigm data (alertness vs fatigue, 32 channels).","marker":"[21]"},{"why":"represents the fixed-channel CNN baselines that DLGE must surpass to handle heterogeneous montages.","marker":"[15]"},{"why":"provides the spatio-spectral modeling of brain regions that motivates the local-global encoding design.","marker":"[10]"},{"why":"Grad-CAM is used to produce the class-activation maps that visualize which brain regions drive classification.","marker":"[25]"}],"fun_headline_variants":["Single EEG model handles three BCI paradigms without retraining","Dual local-global encoder unifies cross-paradigm BCI decoding","Brain-region padding enables one model for three BCI tasks","No retraining needed: one EEG model handles three BCI paradigms","DLGE: one model, three BCI paradigms, ~60% F1"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that random 5-fold cross-validation on shuffled 3-second segments pooled from all three paradigms measures cross-paradigm generalization; because adjacent segments from the same recording are highly similar and no paradigm is ever held out, this premise is what must hold for the reported numbers to support the claim.","fun_headline_variants_meta":{"raw":{"variants":["Single EEG model handles three BCI paradigms without retraining","Dual local-global encoder unifies cross-paradigm BCI decoding","Brain-region padding enables one model for three BCI tasks","No retraining needed: one EEG model handles three BCI paradigms","DLGE: one model, three BCI paradigms, ~60% F1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001185,"raw_usage":{"total_tokens":4766,"prompt_tokens":815,"completion_tokens":3951,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3858}},"tokens_in":559,"tokens_out":3951,"duration_ms":30602,"temperature":1.0,"reasoning_tokens":3858,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:26:17.830624+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DLGE on motor imagery and resting-state data only, then test it on all driving-fatigue data without any fatigue samples in the training folds. If the held-out-paradigm F1 falls to near chance, the central claim fails. As a second check, compare the current shuffled-segment result with a subject-disjoint split; a large drop would indicate that the model exploited within-recording segment similarity.","supporting_citations":[{"cited_title":"Miller, Gernot R","cited_arxiv_id":null,"evidence_quote":"supplies the motor imagery paradigm data (four classes, 22 channels) used in the cross-paradigm evaluation."},{"cited_title":"Prediction of Reaction Time and Vigilance V ar iability From Spatio-Spectral Features of Resting-State EEG in a Lon g Sustained Attention Task","cited_arxiv_id":null,"evidence_quote":"supplies the resting-state paradigm data (eyes open vs closed, 64 channels)."},{"cited_title":"Driver fatigue de tection through multiple entropy fusion analysis in an EEG-based sy stem","cited_arxiv_id":null,"evidence_quote":"supplies the driving fatigue paradigm data (alertness vs fatigue, 32 channels)."},{"cited_title":"EEGNet: a compact convolutional neural network for EEG-based brain–compute r interfaces","cited_arxiv_id":null,"evidence_quote":"represents the fixed-channel CNN baselines that DLGE must surpass to handle heterogeneous montages."},{"cited_title":"Abbass, Kay Chen Tan, Abdullah Al- Mamun, Nitish Thakor, Anastasios Bezerianos, and Junhua Li","cited_arxiv_id":null,"evidence_quote":"provides the spatio-spectral modeling of brain regions that motivates the local-global encoding design."},{"cited_title":"Grad-CAM: Vis ual Explanations From Deep Networks via Gradient-Based Locali zation","cited_arxiv_id":null,"evidence_quote":"Grad-CAM is used to produce the class-activation maps that visualize which brain regions drive classification."}],"review_version":1}