{"id":"eb1dfbb4-dfe2-4d61-b4f8-b63af5f942d7","arxiv_id":"2509.01073","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A LaBraM-based model with IMU attention mapping is fine-tuned for EEG motion artifact removal, but its evaluation metric is identical to its training loss and no independent check of neural signal preservation is provided.","lead":"This paper fine-tunes a large EEG model to remove motion noise from brain recordings by using IMU motion sensors as a reference, and it reports lower EEG-IMU correlation than the ASR+ICA baseline. The catch is that the success metric is the same score the model was trained to minimize, so the reported improvement may be partly by construction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation metric is the training objective and is never validated against ground-truth EEG, so lower coherence scores may reflect blanket signal suppression rather than artifact removal.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the coherence metric is unvalidated as a proxy for artifact removal. My analysis focuses on the circularity that the metric is both the training loss and the evaluation criterion, which makes the reported improvements expected regardless of whether true neural activity is preserved. The suggested concrete test would break this circularity by providing a ground-truth comparison, directly testing whether reduced EEG-IMU correlation corresponds to actual artifact removal rather than signal suppression. The paper also lacks a held-out evaluation and neural-preservation checks, which further weaken the central claim. These are not merely missing baselines; they are essential to support the interpretation of the results. I agree with the reader's REJECT verdict; no adjustment is needed.","tokens_in":8081,"tokens_out":3801,"duration_ms":46575,"concrete_test":"Use the standing-condition EEG from the same Mobile BCI dataset as a clean reference. Synthetically contaminate it by adding a scaled copy of simultaneously recorded IMU accelerometer/gyroscope signals plus small noise, creating paired contaminated/clean EEG. Run the proposed model and ASR+ICA on the contaminated data, then compute (i) RMSE and Pearson correlation between each output and the known clean EEG, and (ii) ERP classification accuracy (target vs. non-target) on the cleaned signals. If the proposed method does not outperform ASR+ICA on RMSE/clean-correlation, or if ERP accuracy drops below the ASR+ICA output, the reduced coherence score is not evidence of artifact removal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the fine-tuned model outperforms ASR+ICA for motion artifact removal rests entirely on the coherence score defined in Section III-B (Eqs. 2-3), which is used both as the training loss and the evaluation metric. Because the model is optimized to minimize exactly this quantity, the lower scores reported in Table I are a direct consequence of optimization, not independent evidence of artifact removal. A degenerate solution—e.g., attenuating all low-frequency EEG content or reducing overall signal amplitude—would also lower the score while destroying neural activity. No ground-truth clean EEG, task-performance metric, or signal-preservation check is provided; the paper explicitly defers comparison with other methods and validation of generalizability to future work (Section V). Additionally, the evaluation uses only 11 participants with no clear train/test split, so the reported numbers may reflect overfitting. Without an independent measure of what the model preserves, the observed coherence reductions cannot support the claim that the method 'disentangles motion artifacts from neural signals' (Section IV-B).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fine-tuned large brain model (LaBraM) with an IMU-based attention mapping module to remove motion artifacts from EEG. Using the Mobile BCI dataset, the authors fine-tune a ~9.2M-parameter model on 5.9 hours of EEG/IMU data from 11 participants, and compare it against an ASR+ICA baseline. The evaluation relies on a weighted frequency-domain EEG–IMU coherence score defined in Eq. (2)–(3). The paper reports that the fine-tuned model consistently achieves the lowest coherence scores across three motion conditions and three time scales.","tokens_in":8333,"tokens_out":3640,"duration_ms":44914,"significance":"If the claimed artifact-removal performance were independently validated, the work would be of interest to the mobile BCI community: it demonstrates parameter-efficient fine-tuning of a large pretrained EEG model (0.2346% of the pretraining data), integrates IMU reference signals through attention, and provides interpretable attention maps. The efficiency and multimodal design are genuine strengths. However, the quantitative evidence currently rests on a metric that is also the training objective, and no ground-truth clean EEG or task-performance validation is provided. The central claim is therefore not supported by the reported experiments as they stand.","major_comments":[{"comment":"The coherence score is explicitly used as both the training loss and the evaluation metric. Section III-B states: 'We use this coherence score as both a quantitative evaluation metric and a training loss signal.' Consequently, the lower coherence values reported for the fine-tuned model in Table I are expected by construction: the model is directly optimized to minimize this quantity. This is a circular evaluation. To support the claim of motion artifact removal, the authors must validate against an independent measure, such as simulated ground-truth clean EEG with known injected motion artifacts, or task performance (e.g., ERP/SSVEP classification accuracy), or neural-signal preservation metrics.","section":"§III-B, Eq. (2)–(3); Table I"},{"comment":"The claim that the learned attention maps 'closely match' the frequency-domain correlation matrix is weakened by the fact that the attention scores are supervised with a scaled version of that same correlation matrix during training (Section III-A: 'a supervision loss that aligns the attention scores (prior to softmax) with a scaled correlation matrix'). Figure 3 therefore shows agreement with a training target, not independent evidence that the model identifies motion-related artifacts. The authors should report attention–correlation agreement on held-out segments or against motion labels not used in the loss.","section":"§IV-A, Fig. 3; §III-A"},{"comment":"No train/test split or cross-subject validation is described. Section II states that the selected ERP recordings from 11 participants were used 'for model training,' while Section IV-B reports correlations on the same three active-motion sessions without clarifying whether these windows overlap with the training data. If the evaluation is performed on training data, the small standard deviations and consistent improvements may reflect overfitting. The authors must specify how evaluation segments were disjoint from training (e.g., held-out subjects, sessions, or time windows) and report cross-subject or cross-session results.","section":"§IV-B, Table I; §II"},{"comment":"The coherence metric itself is not validated as a proxy for artifact removal. The fixed weights (0.7 amplitude, 0.3 phase) and the 0–20 Hz band are chosen without supporting evidence, and a model that globally attenuates low-frequency EEG content or reduces signal amplitude would also lower the coherence score while destroying neural activity. The paper provides no check that the removed components are motion-related rather than neural. At minimum, the authors should report spectral preservation outside the artifact band, ERP peak amplitudes, or classification performance on the cleaned EEG to demonstrate that neural information survives.","section":"§III-B, Eq. (2)"}],"minor_comments":[{"comment":"The notation 'vqnsp encoder base decoder 3x200x12' is unclear; please define the LaBraM variant and the meaning of '3x200x12'.","section":"§III-A"},{"comment":"The text says 'As illustrated in Fig. 2' when referring to comparison results, but Fig. 2 is the framework overview. The intended figure appears to be Fig. 4 or an appendix figure.","section":"§IV-C"},{"comment":"Typo: 'LaBram quantizer layer' should be 'LaBraM quantizer layer'.","section":"§III-A"},{"comment":"Standardizing amplitude and phase before computing Pearson correlation is redundant because Pearson correlation is scale-invariant. Clarify the purpose of this normalization.","section":"§III-B, Eq. (1)"},{"comment":"The session labels 'ses-05', 'ses-04', 'ses-03' are not defined in the text. Specify which recording session each corresponds to and whether these are from the same participants used for training.","section":"Table I"},{"comment":"The manuscript does not explain why only 11 of 24 participants were used or how participants with missing IMU/EEG segments were excluded. Selection criteria should be reported to assess potential bias.","section":"§II"},{"comment":"ASR parameters (e.g., cutoff threshold, reference window) are not reported; only ICLabel thresholds are given. Add ASR settings for reproducibility.","section":"§III-C"},{"comment":"The caption refers to 'two 2-s signal segments,' but the model processes 1-second frames. Clarify how 2-second segments are formed or analyzed.","section":"Fig. 3 caption"}],"recommendation":"reject","confidential_remarks":"This is a well-written preliminary report, but the single quantitative evaluation is circular: the evaluation metric is the training loss. The paper explicitly defers comparisons with other methods and generalizability validation to future work (Section V). Adding the necessary independent validation (ground-truth clean EEG, task performance, or cross-subject splits) would constitute a substantial new study rather than a routine revision. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things. The paper has a genuinely new architecture: fine-tuning LaBraM with an IMU-conditioned attention mapping for EEG motion artifact removal. I have not seen that combination in the literature. And its central evaluation is circular: the coherence score in Eq. (2)-(3) is used as both training loss and the sole reported metric (Section III-B). The paper's own equations make the reported improvements a restatement of the optimization objective.\n\nWhat it does well: the idea of leveraging a pretrained large brain model and aligning IMU features in latent space is sensible, and the attention maps matching the frequency-domain correlation matrices (Fig. 3) is a nice sanity check. The authors are upfront in Section V that they have not compared with other state-of-the-art methods and need more validation. That honesty is worth noting.\n\nThe soft spots are load-bearing rather than minor. No held-out test set is described; Table I appears to report training-condition performance. Only 11 of 24 participants were used, and there is no cross-subject or cross-session evaluation. The coherence metric itself is a proxy: reducing correlation between EEG and IMU in the 0-20 Hz band could be achieved by suppressing broadband EEG activity, not specifically motion artifacts. No ground-truth clean EEG, task-performance metric, or neural-signal-preservation check is provided. Figure 4 is visually suggestive but cannot support the quantitative claim.\n\nSo the paper as written does not support the claim that their method outperforms ASR+ICA. The flaws are fixable: a proper held-out split, validation against a known clean signal or downstream task, and inclusion of the cited IMU baselines (iCanClean, adaptive filtering) would make this a useful contribution.\n\nThis is a paper for a BCI/EEG methods reader who is interested in large-model approaches. It deserves a serious referee, not a desk reject, because the method is plausible and the evaluation can be fixed with a revision. I would not cite the current results, but I might bring it to a reading group as a case study in circular evaluation.\n\nRecommendation: send to peer review, but expect major revision and require independent validation.","headline":"A clever architecture undercuts itself by evaluating on the same correlation metric it optimizes, so the headline claim is unproven.","tokens_in":8820,"tokens_out":2318,"would_cite":false,"duration_ms":25915,"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 fine-tuning a large brain model (LaBraM) with IMU-derived attention removes EEG motion artifacts during walking and running better than ASR+ICA, cutting EEG–IMU correlation by over 50% versus raw EEG.","keywords":["EEG","motion artifact removal","IMU","LaBraM","large brain model","attention mapping","artifact subspace reconstruction","brain-computer interface"],"falsifier":"Record EEG during motion while simultaneously recording a known clean neural signal—for example, use a phantom EEG source or add a simulated movement artifact to clean resting EEG—then check whether the method's lower coherence corresponds to higher signal-to-noise ratio of the known neural events, or instead to broad attenuation of all EEG content. Alternatively, measure ERP or SSVEP classification accuracy before and after cleaning and compare it against ASR+ICA.","tokens_in":7991,"feed_emoji":"🧠","tokens_out":3147,"duration_ms":37263,"temperature":0.7,"pith_summary":"The paper tries to show that EEG motion artifacts can be removed more effectively by combining a pretrained large brain model with inertial measurement unit (IMU) signals than by conventional single-modality methods like ASR and ICA. It fine-tunes LaBraM, a transformer-based EEG encoder, using only 5.9 hours of EEG and IMU recordings, and adds an attention mechanism that aligns EEG features with IMU features in a shared latent space. The claimed result is that this method consistently achieves the lowest EEG–IMU frequency correlation across walking, fast walking, running, and three time scales, beating both raw EEG and the ASR+ICA benchmark. If correct, the work would make mobile brain-computer interfaces more practical by automatically suppressing motion artifacts without manual component selection or task-specific calibration.","feed_headline":"IMU-guided brain model cuts EEG motion artifacts by half","feed_subtitle":"Fine-tuned LaBraM with IMU attention lowers EEG–IMU correlation below ASR+ICA across walking, running, and three time scales.","key_machinery":"The central mechanism is a correlation attention mapping module: EEG signals are encoded by the pretrained LaBraM encoder into 64-dimensional features, IMU signals are projected into the same space by a convolutional encoder, and an attention matrix uses EEG queries and IMU keys to compute channel-wise weights that estimate motion artifact contributions. An artifact gate layer, implemented as an MLP, decides what fraction of the estimated artifact to subtract, and a LaBraM decoder reconstructs clean EEG in the time domain. The whole pipeline is trained and evaluated with the weighted frequency correlation score, which averages Pearson correlations between EEG and IMU amplitude and phase spec","core_discovery":"The paper's central claim is that a fine-tuned LaBraM model with IMU-based correlation attention mapping suppresses EEG motion artifacts better than the established ASR+ICA pipeline. Using a weighted frequency correlation metric over the 0–20 Hz band, the method reduces EEG–IMU correlation by over 50% compared to raw EEG and by over 13% compared to ASR+ICA in every tested condition. The learned attention weights closely match frequency-domain correlation matrices between EEG and IMU channels, which the authors take as evidence that the model learns to identify motion-related EEG content. The fine-tuned model uses only about 0.23% of the training data used to pretrain the base model, suggesti","pith_inferences":["The paper's evaluation metric is also its training loss, so the reported improvements may partly reflect that the model is directly optimized to lower the coherence score; a separate validation against known clean EEG or task performance (e.g., ERP/SSVEP classification accuracy) would be a stronger test of artifact removal.","One testable extension would be to inject simulated motion artifacts into clean resting EEG and check whether the method recovers the known neural signal, rather than merely attenuating all EEG content in the 0–20 Hz band.","The attention-weight alignment with correlation matrices is shown for only a couple of segments; a systematic comparison across all participants and conditions would clarify whether the interpretability is consistent or cherry-picked.","Because the model subtracts estimated motion artifacts in latent space, it may also suppress genuine neural activity that happens to correlate with movement; comparing event-related potentials or steady-state visual evoked potentials before and after cleaning would reveal whether neural information is preserved."],"forward_implications":["If the claim holds, mobile EEG systems could clean motion artifacts automatically during walking, running, and other natural activities, without manual ICA component selection or per-subject calibration.","The result suggests that large pretrained brain models can be fine-tuned for a neural signal-processing task with very little task-specific data—about 0.23% of the pretraining corpus.","The attention maps provide an interpretable link between EEG and IMU channels, which could help diagnose which motion axes most contaminate each EEG electrode.","The method's low correlation scores across 10-second, 30-second, and 1-minute windows imply that the artifact removal is stable across different analysis time scales.","The framework is multimodal in principle, so the same attention-based subtraction idea could be extended to other reference sensors such as electromyography or electrooculography."],"supporting_citations":[{"why":"Supplies the Mobile BCI dataset of scalp EEG and head-mounted IMU recordings during standing, walking, and running, which is the training and evaluation data for the method.","marker":"[10]"},{"why":"Defines LaBraM, the large brain model pretrained on 2500 hours of EEG, whose encoder and decoder are fine-tuned for artifact removal.","marker":"[11]"},{"why":"Provides the Artifact Subspace Reconstruction (ASR) method used as part of the benchmark baseline.","marker":"[3]"},{"why":"Supplies the ICA blind source separation framework that, together with ICLabel classification, forms the other half of the ASR+ICA benchmark.","marker":"[4]"},{"why":"Represents a prior IMU-assisted artifact removal method (iCanClean) that the paper positions as related work and whose limitations motivate the deep-learning approach.","marker":"[9]"},{"why":"Demonstrates adaptive filtering with IMU-derived signals for gait-related EEG artifacts, providing the line of work that the paper extends with a fine-tuned large model.","marker":"[6]"}],"fun_headline_variants":["IMU-aided LaBraM cuts EEG motion artifacts by over 50%","EEG motion artifacts halved using IMU and fine-tuned LaBraM","With IMU cues, tiny fine-tuned LaBraM beats ASR+ICA by 13%","IMU reference signals help fine-tuned LaBraM slash EEG artifacts","EEG+IMU fusion with fine-tuned LaBraM reduces motion artifacts by 50%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing assumption is that a lower EEG–IMU frequency coherence score, averaged over the 0–20 Hz band with fixed amplitude and phase weights, actually means that motion artifacts are removed while neural activity is preserved.","fun_headline_variants_meta":{"raw":{"variants":["IMU-aided LaBraM cuts EEG motion artifacts by over 50%","EEG motion artifacts halved using IMU and fine-tuned LaBraM","With IMU cues, tiny fine-tuned LaBraM beats ASR+ICA by 13%","IMU reference signals help fine-tuned LaBraM slash EEG artifacts","EEG+IMU fusion with fine-tuned LaBraM reduces motion artifacts by 50%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3018,"prompt_tokens":765,"completion_tokens":2253,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":2154}},"tokens_in":509,"tokens_out":2253,"duration_ms":20188,"temperature":1.0,"reasoning_tokens":2154,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:53:03.800986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record EEG during motion while simultaneously recording a known clean neural signal—for example, use a phantom EEG source or add a simulated movement artifact to clean resting EEG—then check whether the method's lower coherence corresponds to higher signal-to-noise ratio of the known neural events, or instead to broad attenuation of all EEG content. Alternatively, measure ERP or SSVEP classification accuracy before and after cleaning and compare it against ASR+ICA.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Mobile BCI dataset of scalp EEG and head-mounted IMU recordings during standing, walking, and running, which is the training and evaluation data for the method."},{"cited_title":"Lee, G.-H","cited_arxiv_id":null,"evidence_quote":"Defines LaBraM, the large brain model pretrained on 2500 hours of EEG, whose encoder and decoder are fine-tuned for artifact removal."},{"cited_title":"Chang, S.-H","cited_arxiv_id":null,"evidence_quote":"Provides the Artifact Subspace Reconstruction (ASR) method used as part of the benchmark baseline."},{"cited_title":"Chang, S.-H","cited_arxiv_id":null,"evidence_quote":"Supplies the ICA blind source separation framework that, together with ICLabel classification, forms the other half of the ASR+ICA benchmark."},{"cited_title":"Lee, N.-S","cited_arxiv_id":null,"evidence_quote":"Represents a prior IMU-assisted artifact removal method (iCanClean) that the paper positions as related work and whose limitations motivate the deep-learning approach."},{"cited_title":"Chuang, K.-Y","cited_arxiv_id":null,"evidence_quote":"Demonstrates adaptive filtering with IMU-derived signals for gait-related EEG artifacts, providing the line of work that the paper extends with a fine-tuned large model."}],"review_version":1}