{"id":"3ed1e685-0f2f-4fac-b4b9-e9ed17530e14","arxiv_id":"2501.01725","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An EEGNet variant with squeeze-and-excitation electrode and filter ranking reaches about 58.8% accuracy in online decoding of left versus right imagined hand movements.","lead":"This paper adds attention-style layers that weight important electrodes and features to an EEGNet deep network for decoding whether a person imagines moving their dominant hand left or right, reaching about 59% accuracy on online EEG sessions. A generalist should care because natural directional control could make brain-computer interfaces for stroke rehabilitation more intuitive, though the gains over existing baselines are small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central accuracy claim may reflect eye-movement artifacts: no EOG rejection, and §III.B itself flags AF7 as noise/eye-related; direction-correlated ocular activity could explain the ~4% advantage.","rationale":"The reader's weakest assumption is the same one I consider load-bearing: ocular artifacts correlated with imagined direction could account for the reported accuracy advantage. The paper itself supplies supporting evidence for this concern in §III.B by ranking AF7 high and attributing it to eye-related artifacts, while the preprocessing pipeline in §II.B has no EOG rejection. If direction-correlated eye movements occur during the 4-s cue, the electrode ranking layer's channel-wise averaging and scaling would allow the model to exploit them directly. I considered alternative concerns: missing statistical significance tests, possible overfitting when fine-tuning on 72 calibration trials, and lack of code/data. These are real but secondary; they reduce confidence without necessarily invalidating the central claim. The artifact confound is the one that, if true, would make the headline comparison meaningless rather than merely imprecise. The recommended verdict remains conditional because the concern is testable and does not require imputing bad faith: a clean EOG/ICA control experiment would either confirm the neural origin of the decoding or demonstrate that the model is reading ocular direction. Until that check is performed, the claim should not be accepted as established.","tokens_in":11804,"tokens_out":4093,"duration_ms":46828,"concrete_test":"Obtain the raw online EEG (or run a replication with synchronized eye tracking or EOG). Estimate horizontal ocular activity from frontal channels Fp1, Fp2, AF7, and AF8 (e.g., via regression or ICA), remove those ocular components from each 4-s MI epoch, then retrain and evaluate the proposed subject-specific model and EEGNet exactly as in §III. If the proposed method's accuracy drops to near chance or its advantage over EEGNet disappears, the claim of neural direction decoding is unsupported. If accuracy remains around 58% and the EEGNet comparison is unchanged, the artifact concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the SE-ranked EEGNet variant decodes direction from unilateral MI at 58.77±8.10% online, beating EEGNet (54.50±5.75%) and FBCNet (49.99±6.73%). The weakest point is an artifact confound. Preprocessing in §II.B is bandpass 0.5–90 Hz, notch, baseline correction, and CAR; there is no EOG rejection. The task is left/right imagined hand movement following a directional visual cue, so direction-correlated saccades or smooth pursuit during the 4-s MI cue can inject frontal ocular potentials into the analyzed window. This is not a speculative edge case: §III.B reports that AF7 is assigned a high rank and attributes it to noise and eye-related artifacts, and no eye-tracker or EOG montage is described. Because the electrode ranking layer (Eq. 1a–1c) averages each channel over the trial and then scales the entire channel, it can directly amplify a low-frequency ocular offset that differs between left and right cues. The observed ~4-point advantage over EEGNet could therefore be artifact classification rather than neural direction decoding. The paper also omits per-subject chance-level tests; with only 48 online trials per subject, some subjects are near or below chance (e.g., S10 at 43.75%). The absence of ocular artifact control is the load-bearing gap: the headline number is only meaningful if it survives removal of ocular components.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep learning architecture for online decoding of left/right directional information from unilateral motor imagery in 20 healthy subjects. The architecture augments EEGNet with Squeeze-and-Excitation (SE) layers that rescale electrodes and feature maps, and the authors evaluate a subject-independent base model and several subject-specific fine-tuned variants on held-out online sessions (13 subjects, 48 trials per subject). The central claim is that the best subject-specific configuration reaches 58.77 ± 8.10% average online accuracy, outperforming EEGNet and FBCNet. The paper also analyzes electrode and filter rankings as an interpretability result.","tokens_in":12055,"tokens_out":5069,"duration_ms":49260,"significance":"The question of decoding directional information from unilateral MI in an online setting is relevant to natural BCI control, and the use of held-out online session data is a strength: the reported means are direct online measurements rather than offline cross-validation. The SE-based ranking mechanism is a simple, transferable idea, and the electrode/filter-rank visualizations offer a useful interpretability tool. However, the central outperformance claim is not yet statistically supported, and the lack of ocular artifact control leaves an unaddressed confound that could explain part of the advantage. If the necessary robustness and significance analyses confirm the effect, the contribution would be a modest but useful step for online directional MI decoding.","major_comments":[{"comment":"The claim that the proposed method outperforms EEGNet and FBCNet is based on group means without any significance testing. With n = 13 subjects, the between-model gaps are within the across-subject variability: for the subject-independent comparison (58.28 ± 9.00 vs. 54.50 ± 5.75), a paired t-test would require per-subject differences, but the reported SDs imply an approximate t-statistic below 2, which is not significant at the 5% level. The subject-specific comparison (58.77 ± 8.10 vs. 55.28 ± 10.05) is likewise inconclusive. Please report per-subject accuracies, paired confidence intervals, and p-values, and state the decision rule for claiming superiority.","section":"Section III.A, Tables I and II"},{"comment":"The preprocessing (0.5–90 Hz bandpass, notch, CAR) does not include EOG rejection, and Section III.B explicitly attributes the high rank of electrode AF7 to noise and eye-related artifacts. Since the electrode ranking layer (Eq. 1a–1c) scales whole channels based on the trial-mean voltage, a low-frequency ocular offset that differs between left and right cues can directly boost classification. The observed ~4-point advantage over EEGNet is therefore potentially artifact classification rather than neural direction decoding. The authors should test whether the result survives removal of frontal/ocular components (e.g., ICA or regression) or provide concurrent EOG evidence.","section":"Section II.B and Section III.B"},{"comment":"With only 48 online trials per subject, the standard error of a single-subject accuracy is about 7.2 percentage points. The observed range (e.g., S10 at 43.75%, S14/S20 at 70.31%) includes many values that are not individually distinguishable from chance. The paper should report binomial chance-level tests per subject and either exclude or separately analyze participants who perform at or below chance, because averaging over such subjects can bias the headlined mean and the model-comparison conclusions.","section":"Section III.A, Fig. 4"}],"minor_comments":[{"comment":"The pooled feature vector in (2a) uses X(i, j) although the input to the feature-map ranking layer is Y; it should be Y(i, j).","section":"Equation (2a)"},{"comment":"Several labels in the architecture diagram are incomplete (e.g., 'on  D', ' atch ormali ation'), making the figure difficult to interpret.","section":"Fig. 2"},{"comment":"The abstract reports 58.7 ± 8% while Table II reports 58.77 ± 8.10%; please make these values consistent.","section":"Abstract vs. Table II"},{"comment":"The filter-ranking observations are based on only three subjects and are post hoc; the text should label them as exploratory rather than as confirmatory evidence.","section":"Section III.C"},{"comment":"Reference [18] (EEGNet) is missing its full citation details, including the journal or conference and year.","section":"Reference [18]"},{"comment":"The term 'Continued Training' is not defined in the text; please specify whether it fine-tunes all layers and with what learning-rate schedule.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core contribution is an online, held-out evaluation of an SE-augmented EEGNet for directional MI decoding. The main barriers are statistical and artifact-related rather than conceptual; both are addressable with additional analyses. The current support for the headline superiority claim is insufficient, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a fairly straightforward integration of SE blocks into EEGNet for left/right imagined hand direction, evaluated on a same-day online session. The new bit is applying SE ranking to both electrodes and feature maps for this task, and reporting online accuracy. That is genuinely new relative to the cited literature, which mostly covers bilateral MI and offline evaluations.\n\nIt does some things right. The evaluation is on actual online session data (13 subjects, 48 trials each), not a replayed offline analysis. The comparison includes feature-engineering baselines like W-CSP and W-PLV, and the subject-specific fine-tuning of only the SE layers is a sensible way to adapt with limited data. The electrode and feature-map ranking analysis is interpretable and biologically plausible: high ranks near motor cortex, low ranks for frontal electrodes.\n\nWhere it gets weak is twofold. First, no statistics anywhere. The reported means have standard deviations, but no paired tests, no confidence intervals, no chance-level comparison. With 48 trials per subject, the 4-point gap over EEGNet could easily be noise; some subjects sit around or below 50%. This is fixable and should be required.\n\nSecond, and more load-bearing, is the eye-artifact issue. The preprocessing is bandpass 0.5–90 Hz, notch, baseline correction, and CAR, with no EOG rejection. The task is left/right imagined movement cued by a directional visual stimulus, so saccades or smooth pursuit during the 4-s MI window are plausible. The paper itself, in Section III.B, says the high rank assigned to AF7 is \"attributed to the noise, as the proposed work did not dwell deep to pre-process the signals of eye-related artifacts.\" That is an admission that ocular signals are in the data. Since the electrode ranking layer averages each channel over the trial and scales the whole channel, it could directly amplify a direction-correlated ocular offset. The ~4% advantage over EEGNet could therefore be artifact classification. The stress-test note is correct on this; it is the central gap.\n\nThat said, the numbers are directly measured, not fit to a prediction, so there is no circularity. The claim is what it is: a possible modest gain, currently unverifiable without code or data and vulnerable to an artifact confound.\n\nWho is this for? BCI researchers working on directional decoding. It is a reasonable conference or workshop paper, but for a journal it needs significance testing plus an artifact control (EOG regression or IC rejection) and ideally data and code. I would send it to review—the question it tackles is real and the online evaluation is a plus—but I would expect major revision. My own verdict would be conditional until the artifact issue is addressed.","headline":"Modest, plausibly real gain for directional MI decoding, undermined by a self-identified ocular artifact confound and no significance testing.","tokens_in":12616,"tokens_out":2232,"would_cite":false,"duration_ms":21136,"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 subject-specific EEGNet variant with Squeeze-and-Excitation ranking layers reaches 58.77% average accuracy for online decoding of motor-imagery direction, beating EEGNet and FBCNet baselines.","keywords":["motor imagery","brain-computer interface","direction decoding","unilateral motor imagery","EEG","Squeeze-and-Excitation","electrode ranking","feature-map ranking"],"falsifier":"Run the same calibration and online protocol with horizontal and vertical EOG channels recorded and removed by regression or rejection, or with subjects instructed to keep gaze fixed during the MI cue, and compare the average accuracy. If the 58.8% advantage over EEGNet collapses toward chance or toward the baseline accuracy, the central claim fails.","tokens_in":11575,"feed_emoji":"🧠","tokens_out":10443,"duration_ms":100417,"temperature":0.7,"pith_summary":"The paper tries to establish that adding Squeeze-and-Excitation (SE) ranking layers to an EEGNet-style convolutional network improves online decoding of left-versus-right direction from unilateral motor imagery in EEG, a task that matters because it could give stroke patients a more natural, higher-dimensional BCI control signal than standard left-hand/right-hand imagery. On data from 20 healthy subjects, with an online session for 13 of them, the paper reports 58.28 ± 9.00% average accuracy for the subject-independent model and 58.77 ± 8.10% for the highest-accuracy subject-specific fine-tuned variant, compared with 54.50 ± 5.75% for EEGNet and 49.99 ± 6.73% for FBCNet. The claimed gain comes from learned ranking layers that reweight electrodes and feature maps, and the same ranking outputs double as an interpretability map. The subject-specific model also shows that only the ranking and dense layers need fine-tuning, not the convolutional feature extractor.","feed_headline":"Electrode ranking lifts imagined-hand direction decoding to 58.8%","feed_subtitle":"Learned electrode and filter importance lifts online motor-imagery direction decoding above standard baselines.","key_machinery":"The central mechanism is the Squeeze-and-Excitation ranking layer applied along two axes. The electrode-ranking layer averages each channel across time, passes the resulting vector through a small dense bottleneck with ReLU and sigmoid activations, and multiplies each electrode's signal by its learned importance weight; the feature-map-ranking layer performs the same operation across the spatial and temporal dimensions of each convolutional filter. The paper fixes the EEGNet convolution filters after pre-training and updates only these ranking layers plus the dense layer for each subject, so the SE weights carry the subject-specific adaptation. The output weights are also the paper's interpretability result: they can be visualized as heatmaps and filter-rank plots to see which electrodes and filters drive the decision.","core_discovery":"On its own terms, the paper's central claim is that SE blocks can be moved from image-channel attention to EEG electrode and feature-map attention, and that this is enough to beat standard deep decoders on directional unilateral MI. The model pre-trains a subject-independent base on calibration data from subjects S01-S07, then for each online subject fine-tunes only the electrode-ranking layer, the feature-map-ranking layer, and the dense layer. The highest-accuracy subject-specific configuration reaches 58.77 ± 8.10% average online accuracy, with the top per-subject accuracy 70.31% (S14 and S20) and the lowest 43.75% (S10). The electrode heatmaps assign high importance to central electrodes near the motor cortex (FC3, FCz, FC4, C3, Cz, C2, CP1, CPz, CP4) and low importance to frontal electrodes such as Fp1, Fp2, AF3, and AFz, which the paper reads as evidence that the ranking is capturing motor-related cortical activity.","pith_inferences":["The authors do not test artifact rejection, and they attribute the high rank of electrode AF7 to eye-related noise; an immediate extension would be to compare accuracy with and without EOG regression, which would isolate how much of the 58.8% is neural direction information.","The same SE-ranking recipe could be dropped into other EEG-BCI paradigms with shift-invariant channel relevance, such as P300 spellers or SSVEP control, where per-user electrode importance could be learned in a few trials.","Their subject-independent model already sits at 58.28%, so a zero-calibration deployment may be feasible; a follow-up needs to check whether the extra 0.5% from fine-tuning justifies the 10-minute calibration session in practice.","A natural next comparison is to run the same subject-independent/subject-specific split against the self-attention unilateral-MI decoder reported in the introduction, since its reported roughly 59% accuracy is the closest point of reference for the proposed method."],"forward_implications":["If the reported accuracy is right, fine-tuning only the ranking and dense layers outperforms continued training of all weights, so the learned importance weights, not the convolutional filters, are what adapt the model to a new subject.","The high ranks assigned to motor-cortex electrodes and low ranks to prefrontal electrodes suggest the ranking layers could serve as a data-driven channel-selection map before a real deployment.","Because the feedback-period EEG was excluded, the online results are attributable to the 4-second motor-imagery window alone, which means the architecture itself, not feedback artifacts, carries the decoding.","For stroke rehabilitation, the framework offers a way to decode left/right intent from a single limb, which would give a BCI more degrees of freedom than conventional bilateral MI without requiring a new feature-engineering pipeline per patient."],"supporting_citations":[{"why":"Supplies the EEGNet architecture whose temporal and spatial convolution filters the proposed model reuses, and provides the main deep-learning baseline in Tables I and II.","marker":"[18]"},{"why":"Introduces the Squeeze-and-Excitation block that the paper extends into electrode and feature-map ranking layers.","marker":"[29]"},{"why":"FBCNet is the second deep-learning baseline; the proposed model is reported to beat its 49.99% average accuracy.","marker":"[19]"},{"why":"Provides the W-CSP feature-engineering baseline for online hand motor-imagery direction decoding that the subject-independent model outperforms.","marker":"[11]"},{"why":"Defines the experimental protocol and the W-CSP plus W-PLV baseline; the same setup supplies the calibration and online session data used here.","marker":"[30]"},{"why":"Establishes the challenge of decoding complex movement intention from overlapping motor-cortex activity, motivating the direction-decoding task.","marker":"[5]"}],"fun_headline_variants":["SE-based electrode ranking sharpens motor imagery direction decoding","58.8% online accuracy for imagined hand direction via electrode ranking","Subject-specific fine-tuning with SE layers improves MI direction decoding","EEG electrode importance from SE layers boosts online MI direction decoding","Learned electrode attention lifts unilateral motor imagery direction decoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results stand on the assumption that the model is learning genuine motor-imagery direction information and not eye-movement artifacts: preprocessing includes bandpass, notch, baseline correction, and common average referencing but no EOG artifact rejection, and the paper itself attributes the high rank of electrode AF7 to noise and eye-related artifacts.","fun_headline_variants_meta":{"raw":{"variants":["SE-based electrode ranking sharpens motor imagery direction decoding","58.8% online accuracy for imagined hand direction via electrode ranking","Subject-specific fine-tuning with SE layers improves MI direction decoding","EEG electrode importance from SE layers boosts online MI direction decoding","Learned electrode attention lifts unilateral motor imagery direction decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001072,"raw_usage":{"total_tokens":4527,"prompt_tokens":1024,"completion_tokens":3503,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":3421}},"tokens_in":640,"tokens_out":3503,"duration_ms":25103,"temperature":1.0,"reasoning_tokens":3421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:38.017244+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same calibration and online protocol with horizontal and vertical EOG channels recorded and removed by regression or rejection, or with subjects instructed to keep gaze fixed during the MI cue, and compare the average accuracy. If the 58.8% advantage over EEGNet collapses toward chance or toward the baseline accuracy, the central claim fails.","supporting_citations":[{"cited_title":"et Compact Convolutional Neural Network for EEG -based Brain- omputer Interfaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the EEGNet architecture whose temporal and spatial convolution filters the proposed model reuses, and provides the main deep-learning baseline in Tables I and II."},{"cited_title":"Squee e -and-Excitation etworks,","cited_arxiv_id":null,"evidence_quote":"Introduces the Squeeze-and-Excitation block that the paper extends into electrode and feature-map ranking layers."},{"cited_title":"F et: a Multi-view CNN with novel Variance Layer for Motor Imagery rain omputer Interface,","cited_arxiv_id":null,"evidence_quote":"FBCNet is the second deep-learning baseline; the proposed model is reported to beat its 49.99% average accuracy."},{"cited_title":"From classic motor imagery to comple mo ement intention decoding: The noninvasive Graz -BCI approach,","cited_arxiv_id":null,"evidence_quote":"Establishes the challenge of decoding complex movement intention from overlapping motor-cortex activity, motivating the direction-decoding task."}],"review_version":1}