Pith. sign in

REVIEW 4 major objections 8 minor 12 references

IMU-Enhanced EEG Motion Artifact Removal with Fine-Tuned Large Brain Models

T0 review · 4 major / 8 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict A clever architecture undercuts itself by evaluating on the same correlation metric it optimizes, so the headline claim is unproven. read the letter →

arxiv 2509.01073 v1 pith:YC7BDERP submitted 2025-09-01 cs.LG eess.SPq-bio.NC

Yuhong Zhang , Xusheng Zhu , Yuchen Xu , ChiaEn Lu , Hsinyu Shih , Gert Cauwenberghs , Tzyy-Ping Jung This is my paper · ORCID
classification cs.LGeess.SPq-bio.NC
keywords EEGmotionartifactremovalIMULaBraMlargebrainmodelattentionmappingsubspacereconstructionbrain-computerinterface
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

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.

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 (4)
  1. [§III-B, Eq. (2)–(3); Table I] 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.
  2. [§IV-A, Fig. 3; §III-A] 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.
  3. [§IV-B, Table I; §II] 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.
  4. [§III-B, Eq. (2)] 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.
minor comments (8)
  1. [§III-A] The notation 'vqnsp encoder base decoder 3x200x12' is unclear; please define the LaBraM variant and the meaning of '3x200x12'.
  2. [§IV-C] 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.
  3. [§III-A] Typo: 'LaBram quantizer layer' should be 'LaBraM quantizer layer'.
  4. [§III-B, Eq. (1)] Standardizing amplitude and phase before computing Pearson correlation is redundant because Pearson correlation is scale-invariant. Clarify the purpose of this normalization.
  5. [Table I] 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.
  6. [§II] 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.
  7. [§III-C] ASR parameters (e.g., cutoff threshold, reference window) are not reported; only ICLabel thresholds are given. Add ASR settings for reproducibility.
  8. [Fig. 3 caption] 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.

Circularity Check

2 steps flagged · score 6.0 of 10

Evaluation metric is the training objective, and the attention-vs-correlation agreement is supervised into the model: the reported coherence reductions are by construction.

  1. fitted input called prediction [Section III-B (Eqs. 2-3) and Section IV-B (Table I)]
    "We use this coherence score as both a quantitative evaluation metric and a training loss signal while fine-tuning our IMU-integrated denoising model. ... In contrast, our fine-tuned LaBraM model consistently achieves the lowest correlation scores across all scenarios."

    Coherence_p in Eq. (3) is the training loss minimized during fine-tuning, and Table I reports exactly this same scalar as the evaluation result. The model is gradient-optimized to reduce this quantity, so the 'prediction' of lower coherence is the fitted objective itself, not an independent measure of artifact removal. Because lower coherence can also be achieved by attenuating all low-frequency EEG content or shrinking signal amplitude, the reported reductions do not by themselves demonstrate that neural activity was preserved. No ground-truth clean EEG or task-performance check is provided to break the equivalence between optimizing the loss and 'validating' the loss.

  2. self definitional [Section III-A and Section IV-A (Fig. 3)]
    "We introduced a supervision loss that aligns the attention scores (prior to softmax) with a scaled correlation matrix, computed by subtracting 0.5 and multiplying by 20. ... The learned attention weights closely mirror the reference correlation patterns, showing clear structural similarity."

    The attention maps are trained to match a scaled version of the same frequency-domain correlation matrix used for evaluation. The Figure 3 agreement is therefore enforced by the supervision loss, not discovered by the model. Presenting this agreement as evidence that the model 'effectively learns to identify and emphasize motion-related dependencies' is circular: the correspondence between attention and correlation is part of the training objective, so it cannot independently validate the attention mechanism or the artifact-removal claim.

full rationale

The paper's central claim—that the fine-tuned LaBraM model achieves superior motion artifact removal—rests on the weighted frequency coherence score defined in Section III-B. The paper explicitly states this score is used 'as both a quantitative evaluation metric and a training loss signal.' Consequently, the Table I reductions are the values of the training objective itself, evaluated without any described held-out split or independent ground-truth signal. Degenerate solutions such as suppressing broad EEG content would lower the same score while destroying neural information, and the paper provides no clean-EEG reconstruction check, task-performance metric, or preservation-of-neural-activity measure to rule this out. The attention-map validation is similarly self-confirming: Section III-A introduces a supervision loss that aligns attention scores with the same correlation matrix, so the Section IV-A agreement in Fig. 3 is constructed rather than emergent. These are not cases of self-citation or imported uniqueness; the LaBraM backbone and dataset are external. The circularity is instead in the evaluation protocol, but it is load-bearing: the paper's stated conclusion that the method 'disentangles motion artifacts from neural signals' is derived from the metric the model was optimized to minimize. Section V further acknowledges that comparisons with other state-of-the-art methods and generalizability validation are deferred. Overall score 6: partial but significant circularity, since the central performance evidence reduces by construction to the training loss, while the method still has some independent architectural content.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the artifact-cleanliness metric being valid, plus the hand-chosen weights in the metric itself. The model architecture contributes no new physical entities. The metric's dual use as loss and evaluation is the primary analytical weakness, and the free parameters (0.7/0.3, attention scaling, band choice) are all fitted or chosen without independent benchmarks.

free parameters (6)
  • amplitude correlation weight = 0.7
    In Eq. (2), amplitude correlation is weighted 0.7 with no justification beyond 'Amplitude is more strongly affected by motion artifacts'; this weight directly shapes the loss and metric.
  • phase correlation weight = 0.3
    In Eq. (2), phase correlation is weighted 0.3; chosen by hand and not varied or justified with evidence.
  • attention scaling offset = 0.5
    In Section III-A, the correlation matrix is transformed by subtracting 0.5 and multiplying by 20 to supervise attention weights; these constants are ad hoc choices.
  • attention scaling factor = 20
    Same as above, the multiplication factor 20 is introduced without derivation or sensitivity analysis.
  • frequency band for coherence = 0-20 Hz (40 bins)
    The 0-20 Hz band is asserted to contain motion artifacts; the boundary and bin count are not derived from data or compared with alternatives.
  • embedding dimension = 64
    Both EEG and IMU embeddings are set to 64 dimensions; this architectural number is chosen without ablation.
assumptions (3)
  • domain assumption EEG motion artifacts primarily occupy the 0-20 Hz band and are captured by frequency-domain linear correlation with IMU signals.
    Invoked in Section III-B where the coherence metric is defined; no empirical or cited support for the band limits is provided.
  • ad hoc to paper Reducing the EEG-IMU correlation score is equivalent to removing motion artifacts while preserving brain activity.
    This is the load-bearing evaluative assumption. The paper uses the score as both loss and metric and never validates it against ground-truth clean EEG or task accuracy.
  • standard math Standard signal processing properties of FFT, Pearson correlation, and layer-wise normalization hold as used.
    No novel mathematics; these are background operations in Eq. (1)-(3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMU-Enhanced EEG Motion Artifact Removal with Fine-Tuned Large Brain Models." pith.science (2026). https://pith.science/paper/YC7BDERP

@misc{pith2026250901073,
  author       = {Pith},
  title        = {Pith review of: IMU-Enhanced EEG Motion Artifact Removal with Fine-Tuned Large Brain Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YC7BDERP}},
  note         = {Machine review of arXiv:2509.01073}
}
read the original abstract

Electroencephalography (EEG) is a non-invasive method for measuring brain activity with high temporal resolution; however, EEG signals often exhibit low signal-to-noise ratios because of contamination from physiological and environmental artifacts. One of the major challenges hindering the real-world deployment of brain-computer interfaces (BCIs) involves the frequent occurrence of motion-related EEG artifacts. Most prior studies on EEG motion artifact removal rely on single-modality approaches, such as Artifact Subspace Reconstruction (ASR) and Independent Component Analysis (ICA), without incorporating simultaneously recorded modalities like inertial measurement units (IMUs), which directly capture the extent and dynamics of motion. This work proposes a fine-tuned large brain model (LaBraM)-based correlation attention mapping method that leverages spatial channel relationships in IMU data to identify motion-related artifacts in EEG signals. The fine-tuned model contains approximately 9.2 million parameters and uses 5.9 hours of EEG and IMU recordings for training, just 0.2346\% of the 2500 hours used to train the base model. We compare our results against the established ASR-ICA benchmark across varying time scales and motion activities, showing that incorporating IMU reference signals significantly improves robustness under diverse motion scenarios.

Figures

Figures reproduced from arXiv: 2509.01073 by the authors.

Figure 1
Figure 1. The experimental setup used in the dataset [10]. For [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed IMU-integrated motion artifact removal framework, which fine-tunes a LaBraM-based [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of EEG–IMU attention maps and their correspondence with correlation matrices across two 2-s signal [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: EEG denoising comparison during fast walking (1.6 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 1
Figure 1. Figure 1: Slight Running (2.0 m/s), 10-second window. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png]
Figure 2
Figure 2. Figure 2: Slight Running (2.0 m/s), 30-second window. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png]
Figure 3
Figure 3. Figure 3: Slight Running (2.0 m/s), 60-second window. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png]
Figure 4
Figure 4. Figure 4: Fast Walking (1.6 m/s), 10-second window. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Fast Walking (1.6 m/s), 30-second window. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Fast Walking (1.6 m/s), 60-second window. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Slow Walking (0.8 m/s), 10-second window. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Slow Walking (0.8 m/s), 30-second window. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Slow Walking (0.8 m/s), 60-second window. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: ICLabel-based classification of ICA components during slight running. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

  1. [1]

    11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

    work page arXiv 2019

  2. [2]

    Mullen, C

    T. Mullen, C. Kothe, Y. M. Chi, A. Ojeda, T. Kerth, S. Makeig, G. Cauwenberghs, and T.-P. Jung, ``Real-time modeling and 3d visualization of source dynamics and connectivity using wearable eeg,'' in 2013 35th annual international conference of the IEEE engineering in medicine and biology society (EMBC) , pp. 2184--2187, IEEE, 2013

    work page 2013

  3. [3]

    Chang, S.-H

    C.-Y. Chang, S.-H. Hsu, L. Pion-Tonachini, and T.-P. Jung, ``Evaluation of artifact subspace reconstruction for automatic eeg artifact removal,'' in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , pp. 1242--1245, IEEE, 2018

    work page 2018

  4. [4]

    Chang, S.-H

    C.-Y. Chang, S.-H. Hsu, L. Pion-Tonachini, and T.-P. Jung, ``Evaluation of artifact subspace reconstruction for automatic artifact components removal in multi-channel eeg recordings,'' IEEE transactions on biomedical engineering , vol. 67, no. 4, pp. 1114--1121, 2019

    work page 2019

  5. [5]

    T.-P. Jung, S. Makeig, C. Humphries, T.-W. Lee, M. J. Mckeown, V. Iragui, and T. J. Sejnowski, ``Removing electroencephalographic artifacts by blind source separation,'' Psychophysiology , vol. 37, no. 2, pp. 163--178, 2000

    work page 2000

  6. [6]

    Chuang, K.-Y

    C.-H. Chuang, K.-Y. Chang, C.-S. Huang, and T.-P. Jung, ``Ic-u-net: a u-net-based denoising autoencoder using mixtures of independent components for automatic eeg artifact removal,'' NeuroImage , vol. 263, p. 119586, 2022

    work page 2022

  7. [7]

    Kilicarslan and J

    A. Kilicarslan and J. L. C. Vidal, ``Characterization and real-time removal of motion artifacts from eeg signals,'' Journal of neural engineering , vol. 16, no. 5, p. 056027, 2019

    work page 2019

  8. [8]

    Beach, M

    C. Beach, M. Li, E. Balaban, and A. J. Casson, ``Motion artefact removal in electroencephalography and electrocardiography by using multichannel inertial measurement units and adaptive filtering,'' Healthcare Technology Letters , vol. 8, no. 5, pp. 128--138, 2021

    work page 2021

Show all 12 references
  1. [9]

    Lee, N.-S

    Y.-E. Lee, N.-S. Kwak, and S.-W. Lee, ``A real-time movement artifact removal method for ambulatory brain-computer interfaces,'' IEEE Transactions on Neural Systems and Rehabilitation Engineering , vol. 28, no. 12, pp. 2660--2670, 2020

    2020

  2. [10]

    R. J. Downey and D. P. Ferris, ``icanclean removes motion, muscle, eye, and line-noise artifacts from phantom eeg,'' Sensors , vol. 23, no. 19, p. 8214, 2023

    2023

  3. [11]

    Lee, G.-H

    Y.-E. Lee, G.-H. Shin, M. Lee, and S.-W. Lee, ``Mobile bci dataset of scalp-and ear-eegs with erp and ssvep paradigms while standing, walking, and running,'' Scientific data , vol. 8, no. 1, p. 315, 2021

    2021

  4. [12]

    Jiang, L.-M

    W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, ``Large brain model for learning generic representations with tremendous eeg data in bci,'' arXiv preprint arXiv:2405.18765 , 2024

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.