arxiv: 2604.24428 · v2 · submitted 2026-04-27 · 📡 eess.SP · cs.AI

Recognition: no theorem link

BandRouteNet: An Adaptive Band Routing Neural Network for EEG Artifact Removal

Phat Lam

Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:18 UTC · model grok-4.3

classification 📡 eess.SP cs.AI

keywords EEG artifact removalneural network denoisingband routingfrequency-aware processingsignal reconstructionEOG EMG artifactsparameter-efficient model

0 comments

The pith

BandRouteNet removes EEG artifacts more effectively than prior methods by routing denoising adaptively across frequency bands while using a full-band conditioner for context and only 0.2 million parameters.

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

The paper introduces BandRouteNet as a neural network that splits EEG processing into separate frequency bands to target the distinct spectral patterns of different artifacts like eye movements and muscle activity. Within each band it applies an adaptive routing step that decides at each moment how much denoising to perform. A parallel full-band pathway extracts overall temporal context from the raw signal and uses that to adjust the band pathways while also adding a coarse cleanup to the final output. Experiments on the EEGDenoiseNet benchmark show this combination yields lower relative error and higher noise improvement scores than other approaches for EOG, EMG, and mixed artifacts. The design keeps the total trainable parameters at 0.2 million, which supports use on devices with limited resources.

Core claim

BandRouteNet performs band-wise denoising to explicitly capture frequency-dependent artifact patterns. Within this framework, a routing mechanism adaptively determines where and to what extent denoising should be applied across temporal locations within each frequency band. In parallel, a full-band conditioner directly processes the original noisy EEG to extract global temporal context, producing both conditional parameters for modulating the band-wise pathway and a coarse-grained signal-level refinement to supplement the final reconstruction.

What carries the argument

The adaptive routing mechanism that selects the location and strength of denoising inside each frequency band, guided by conditional parameters from a full-band conditioner.

If this is right

The approach can handle temporally varying and spectrally distinct artifacts within a single unified model.
Resource-constrained devices such as portable brain-computer interfaces can run high-performance denoising locally.
Unified experimental settings allow direct comparison showing consistent gains across EOG, EMG, and mixed conditions.
Band-specific processing plus global conditioning together produce both fine detail cleanup and overall signal refinement.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same routing idea could be tested on other noisy biosignals such as ECG or MEG to see whether frequency-aware adaptation transfers.
Low parameter count opens the possibility of on-device fine-tuning for individual users in clinical or consumer settings.
If the routing decisions prove stable, the architecture might serve as a modular building block for multi-task EEG pipelines that include both denoising and feature extraction.

Load-bearing premise

The assumption that the band-wise routing and full-band conditioner will generalize beyond the EEGDenoiseNet benchmark and the specific artifact types tested there.

What would settle it

Running the trained model on an independent EEG artifact dataset recorded under different conditions and checking whether the reported gains in Relative Root Mean Square Error and Signal-to-Noise Ratio Improvement remain larger than those of competing methods.

Figures

Figures reproduced from arXiv: 2604.24428 by Phat Lam.

**Figure 1.** Figure 1: Overview of the proposed BandRouteNet architecture for EEG denoising. The left subfigure illustrates the overall framework, while the right subfigure presents the detailed structures of the main component blocks. The model consists of two key modules: a Band-specific Denoiser, which removes artifacts within each decomposed EEG band, and a Full-band Conditioner, which processes the original full-band signal… view at source ↗

**Figure 2.** Figure 2: Denoising performance comparison on the EOG-contaminated seg view at source ↗

**Figure 3.** Figure 3: Denoising performance comparison on the EMG-contaminated seg view at source ↗

**Figure 5.** Figure 5: Performance Metrics on different SNR Levels (EOG Dataset) view at source ↗

**Figure 6.** Figure 6: Performance Metrics on different SNR Levels (EMG Dataset) view at source ↗

**Figure 7.** Figure 7: Performance Metrics on different SNR Levels (Mixed EOG/EEG view at source ↗

**Figure 8.** Figure 8: Artifact routing visualization results under three noise conditions. In each sub-figure, the upper panels present the denoising results, while the lower view at source ↗

read the original abstract

Electroencephalography (EEG) is highly susceptible to artifact contamination, such as electrooculographic (EOG) and electromyographic (EMG) interference, which severely degrades signal quality and hinders reliable interpretation in applications including neurological diagnosis, brain-computer interfaces (BCIs), etc. Effective EEG denoising remains challenging because different artifact sources exhibit diverse and temporally varying distributions, together with distinct spectral characteristics across frequency bands. To address these issues, we propose BandRouteNet, an adaptive frequency-aware neural network for EEG denoising that jointly exploits band-specific processing and full-band contextual modeling. The proposed model performs band-wise denoising to explicitly capture frequency-dependent artifact patterns. Within this framework, we introduce a routing mechanism that adaptively determines where and to what extent denoising should be applied across temporal locations within each frequency band. In parallel, a full-band conditioner directly processes the original noisy EEG to extract global temporal context, producing both conditional parameters for modulating the band-wise pathway and a coarse-grained signal-level refinement to supplement the final reconstruction. Extensive experiments on the EEGDenoiseNet benchmark dataset demonstrate that BandRouteNet outperforms other methods under EOG, EMG, and mixed-artifact conditions in terms of Relative Root Mean Square Error (RRMSE) and Signal-to-Noise Ratio Improvement (SNR$_{\text{imp}}$) under unified experimental settings, while remaining highly parameter-efficient with only 0.2M trainable parameters. These results highlight its strong potential for high-performance EEG artifact removal in resource-constrained applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BandRouteNet adds band-wise adaptive routing plus a full-band conditioner for EEG denoising and reports solid benchmark numbers with low parameter count, but the routing's specific contribution is not clearly isolated.

read the letter

The paper's main contribution is BandRouteNet, which processes EEG signals band by band with an adaptive routing step that decides denoising strength per time location in each band, while a separate full-band path supplies context and helps refine the output. It shows better performance than baselines on the EEGDenoiseNet dataset for different artifact types, using only 0.2M parameters. This architecture is new in how it combines explicit band routing with the conditioner. Prior EEG denoisers tend to treat the signal more uniformly or without the adaptive per-band decision. The work does well on the efficiency side, which is practical for embedded or real-time BCI uses. The results under unified settings are a plus if the comparisons are fair. The soft spots are around the routing mechanism. The claim that it handles diverse and varying artifact distributions rests on this adaptive part, but the paper does not provide ablations showing its isolated effect or details on how the routing is learned and supervised. Without that, it's possible the gains come from the full-band path or other design choices instead. The experiments are on one benchmark, so generalization is an open question. Math and data look standard for this area—no fancy derivations, just empirical NN results. Citations seem to cover the relevant prior denoising methods. This paper is for researchers focused on applied signal processing for brain signals, especially those needing lightweight models. A reader working on BCI artifact removal would find the architecture and numbers worth looking at. I would recommend sending it for peer review. The idea is solid enough to warrant feedback, even though the authors should strengthen the evidence for the routing's role.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes BandRouteNet, an adaptive frequency-aware neural network for EEG artifact removal. It jointly uses band-wise denoising with a routing mechanism that adaptively determines where and to what extent denoising is applied across temporal locations within each frequency band, plus a full-band conditioner that processes the original noisy EEG to extract global context, produce conditional parameters, and provide coarse signal refinement. Extensive experiments on the EEGDenoiseNet benchmark are reported to show outperformance over other methods for EOG, EMG, and mixed-artifact conditions in RRMSE and SNR_imp under unified settings, with only 0.2M trainable parameters.

Significance. If the empirical outperformance holds under full experimental details and the adaptive routing is validated as the key driver via ablations, the work could provide a practical, parameter-efficient approach to EEG denoising suitable for resource-constrained applications such as BCIs. The focus on frequency-band-specific processing and global conditioning addresses a real challenge in artifact removal, and the use of a public benchmark is a positive element.

major comments (3)

[Method section (routing mechanism)] The routing mechanism (described in the method section): no equations, pseudocode, or training details are provided for how adaptive routing decisions are learned or computed (e.g., learned gates, attention, Gumbel sampling, or auxiliary losses). This is load-bearing because the paper positions the routing as the key innovation for handling diverse and temporally varying artifact distributions, yet gains could instead arise from the full-band path or standard CNN blocks.
[Experiments section] Experiments section: the central outperformance claim on EEGDenoiseNet lacks any mention of data splits, number of runs, error bars, statistical tests, or exact baseline implementations. Without these, the assertion of superiority 'under unified experimental settings' cannot be verified and the results remain un-reproducible from the provided information.
[Experiments section] No ablation studies are presented to isolate the contribution of the band-wise routing versus the full-band conditioner (or versus a non-routed band-wise baseline). This is required to substantiate that the advertised adaptive mechanism, rather than other architectural choices or training protocol, drives the reported RRMSE and SNR_imp gains.

minor comments (1)

[Abstract] The abstract refers to 'unified experimental settings' without defining them; a brief clarification of the common protocol (e.g., preprocessing, train/test split) would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough and constructive review. We appreciate the recognition of the work's potential significance for practical EEG denoising applications. Below we address each major comment in detail and commit to a major revision that incorporates all requested clarifications and additional experiments.

read point-by-point responses

Referee: [Method section (routing mechanism)] The routing mechanism (described in the method section): no equations, pseudocode, or training details are provided for how adaptive routing decisions are learned or computed (e.g., learned gates, attention, Gumbel sampling, or auxiliary losses). This is load-bearing because the paper positions the routing as the key innovation for handling diverse and temporally varying artifact distributions, yet gains could instead arise from the full-band path or standard CNN blocks.

Authors: We agree that the current method section lacks sufficient formalization of the routing mechanism. In the revised manuscript we will add the complete mathematical formulation (including the gating function, routing weights, and how they modulate the band-wise denoisers), pseudocode for the forward pass, and explicit training details such as the differentiable routing implementation (via soft attention-style gates) and any auxiliary objectives used to encourage adaptive behavior. These additions will make clear that the routing is the primary driver of handling temporally varying artifacts rather than the full-band path alone. revision: yes
Referee: [Experiments section] Experiments section: the central outperformance claim on EEGDenoiseNet lacks any mention of data splits, number of runs, error bars, statistical tests, or exact baseline implementations. Without these, the assertion of superiority 'under unified experimental settings' cannot be verified and the results remain un-reproducible from the provided information.

Authors: We acknowledge that the experiments section is currently missing these critical reproducibility elements. The revised version will specify the exact train/validation/test splits of EEGDenoiseNet, the number of independent training runs (with different random seeds), standard-deviation error bars on all reported RRMSE and SNR_imp metrics, results of statistical significance tests (e.g., paired t-tests or Wilcoxon tests against baselines), and precise implementation details for each baseline (including hyper-parameters, training protocols, and any re-implementation choices) to enable full verification under unified settings. revision: yes
Referee: [Experiments section] No ablation studies are presented to isolate the contribution of the band-wise routing versus the full-band conditioner (or versus a non-routed band-wise baseline). This is required to substantiate that the advertised adaptive mechanism, rather than other architectural choices or training protocol, drives the reported RRMSE and SNR_imp gains.

Authors: We will add a dedicated ablation subsection in the revised manuscript. This will include quantitative comparisons of: (i) the full BandRouteNet, (ii) a version without the routing mechanism (fixed uniform routing), (iii) a version without the full-band conditioner, and (iv) a non-routed band-wise CNN baseline. The results will directly quantify the incremental gains attributable to the adaptive routing component on both EOG and EMG conditions, thereby confirming its role in the observed performance improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark evaluation is independent of model definition

full rationale

The paper introduces BandRouteNet as a neural architecture for EEG denoising and supports its claims solely through empirical comparisons on the external EEGDenoiseNet benchmark under unified settings. No mathematical derivations, first-principles predictions, or fitted parameters are presented that reduce to the model's own inputs by construction. The routing mechanism and conditioner are architectural choices whose contribution is assessed via external metrics (RRMSE, SNR_imp), not by re-using fitted values or self-citations as load-bearing justification. The evaluation protocol is falsifiable outside the paper and does not invoke uniqueness theorems or ansatzes from prior self-work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the neural architecture on the EEGDenoiseNet dataset; no explicit free parameters, axioms, or invented physical entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5562 in / 1144 out tokens · 37355 ms · 2026-05-12T03:18:00.875626+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 2 internal anchors

[1]

Recent applications of eeg-based brain-computer-interface in the medical field,

X.-Y . Liu et al., “Recent applications of eeg-based brain-computer-interface in the medical field,”Military Medical Research, vol. 12, no. 1, p. 14, 2025

work page 2025
[2]

Identification and removal of physiological artifacts from electroencephalogram signals: A review,

M. M. N. Mannan, M. A. Kamran, and M. Y . Jeong, “Identification and removal of physiological artifacts from electroencephalogram signals: A review,”Ieee Access, vol. 6, pp. 30 630–30 652, 2018

work page 2018
[3]

Ocular reduction in eeg signals based on adaptive filtering, regression and blind source separation,

S. Romero et al., “Ocular reduction in eeg signals based on adaptive filtering, regression and blind source separation,”Annals of biomedical engineering, vol. 37, no. 1, pp. 176–191, 2009

work page 2009
[4]

Online artifact removal for brain- computer interfaces using support vector machines and blind source separation,

S. Halder et al., “Online artifact removal for brain- computer interfaces using support vector machines and blind source separation,”Computational intelligence and neuroscience, vol. 2007, no. 1, p. 082 069, 2007

work page 2007
[5]

Electromyogram (emg) removal by adding sources of emg (erase)—a novel ica-based algorithm for removing myoelectric artifacts from eeg,

Y . Li et al., “Electromyogram (emg) removal by adding sources of emg (erase)—a novel ica-based algorithm for removing myoelectric artifacts from eeg,”Frontiers in neuroscience, vol. 14, p. 597 941, 2021

work page 2021
[6]

Eeg artifact re- moval system for depression using a hybrid denoising approach,

C. Kaur, P. Singh, and S. Sahni, “Eeg artifact re- moval system for depression using a hybrid denoising approach,”Basic and Clinical Neuroscience, vol. 12, no. 4, p. 465, 2021

work page 2021
[7]

Single-channel eeg signal extraction based on dwt, ceemdan, and ica method,

Q. Hu et al., “Single-channel eeg signal extraction based on dwt, ceemdan, and ica method,”Frontiers in Human Neuroscience, vol. 16, p. 1 010 760, 2022

work page 2022
[8]

Zhang et al.,Eegdenoisenet: A benchmark dataset for end-to-end deep learning solutions of eeg denoising,

H. Zhang et al.,Eegdenoisenet: A benchmark dataset for end-to-end deep learning solutions of eeg denoising,

work page
[9]

arXiv: 2009.11662[eess.SP]

work page arXiv 2009
[10]

Eegdnet: Fusing non-local and local self-similarity for eeg signal denoising with trans- former,

X. Pu et al., “Eegdnet: Fusing non-local and local self-similarity for eeg signal denoising with trans- former,”Computers in Biology and Medicine, vol. 151, p. 106 248, 2022,ISSN: 0010-4825

work page 2022
[11]

Embedding decomposition for artifacts removal in eeg signals,

J. Yu et al., “Embedding decomposition for artifacts removal in eeg signals,”Journal of Neural Engineering, vol. 19, no. 2, p. 026 052, Apr. 2022

work page 2022
[12]

Lrr-unet: A deep unfolding network with low-rank recovery for eeg signal denoising,

X. Yue et al., “Lrr-unet: A deep unfolding network with low-rank recovery for eeg signal denoising,”CNS Neuroscience & Therapeutics, vol. 31, no. 10, e70632, 2025

work page 2025
[13]

Denoising eeg signals for real- world bci applications using gans,

E. Brophy et al., “Denoising eeg signals for real- world bci applications using gans,”Frontiers in Neu- roergonomics, vol. V olume 2 - 2021, 2022,ISSN: 2673- 6195.DOI: 10.3389/fnrgo.2021.805573

work page doi:10.3389/fnrgo.2021.805573 2021
[14]

Eeg signal denoising using pix2pix gan: Enhancing neurological data analysis,

H. Wang et al., “Eeg signal denoising using pix2pix gan: Enhancing neurological data analysis,”arXiv preprint arXiv:2411.13288, 2024

work page arXiv 2024
[15]

Removal of ocular artifact from the eeg: A review,

R. J. Croft et al., “Removal of ocular artifact from the eeg: A review,”Neurophysiologie Clinique/Clinical Neurophysiology, vol. 30, no. 1, pp. 5–19, 2000

work page 2000
[16]

Emg contamination of eeg: Spectral and topographical characteristics,

I. I. Goncharova et al., “Emg contamination of eeg: Spectral and topographical characteristics,”Clinical neurophysiology, vol. 114, no. 9, pp. 1580–1593, 2003

work page 2003
[17]

A comprehensive review of deep learning models for denoising eeg signals: Challenges, advances, and future directions,

V . A. Raj et al., “A comprehensive review of deep learning models for denoising eeg signals: Challenges, advances, and future directions,”Discover Applied Sci- ences, vol. 7, no. 11, p. 1268, 2025

work page 2025
[18]

Film: Visual reasoning with a general conditioning layer,

E. Perez et al., “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI confer- ence on artificial intelligence, vol. 32, 2018

work page 2018
[19]

Attention Is All You Need

A. Vaswani et al.,Attention is all you need, 2023. arXiv: 1706.03762[cs.CL]

work page internal anchor Pith review Pith/arXiv arXiv 2023
[20]

A novel end-to-end 1d-rescnn model to remove artifact from eeg signals,

W. Sun et al., “A novel end-to-end 1d-rescnn model to remove artifact from eeg signals,”Neurocomputing, vol. 404, pp. 108–121, 2020,ISSN: 0925-2312

work page 2020
[21]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter,Decoupled weight decay regularization, 2019. arXiv: 1711.05101[cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2019