BandRouteNet: An Adaptive Band Routing Neural Network for EEG Artifact Removal
arXiv:2604.24428v1 [eess.SP] 27 Apr 2026
Phat Lam
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 (SNRimp ) 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. Keywords— EEG artifact removal, Electrooculography (EOG), Electromyography (EMG), EEGDenoiseNet, Routing Mechanism, Band-specific denoising, Full-band conditioner.
I. I NTRODUCTION Electroencephalography (EEG) is a widely used noninvasive tool for monitoring brain activity due to its high temporal resolution, low cost, and portability. It supports a wide range of applications, including brain–computer interfaces (BCIs), neurological monitoring, sleep analysis, and clinical diagnosis [1]. However, EEG recordings are highly susceptible to contamination, particularly from electrooculographic (EOG) and electromyographic (EMG) artifacts, which can severely distort signal morphology, spectral content, and downstream interpretability [2]. A large body of work has focused on EEG denoising. Classical approaches mainly rely on explicit assumptions about artifact generation or signal structure. For example, Regression and adaptive filtering methods use reference channels, such as EOG, to estimate and subtract artifacts, but their performance depends strongly on the availability and quality of these references. [3]. Blind source separation methods particularly Independent Component Analysis (ICA), have been P. Lam is with Ho Chi Minh City University of Technology, Vietnam.
widely used for ocular and muscular artifact removal; however, they often require multichannel recordings, component selection heuristics [4], [5]. Time–frequency and decompositionbased approaches, such as Wavelet Transforms and Empirical Mode Decomposition (EMD), offer greater flexibility for nonstationary EEG signals, but they are sensitive to threshold selection, mode mixing, and decomposition instability [6], [7]. More recently, deep learning has emerged as a promising alternative for EEG denoising, as it can learn a direct mapping from contaminated EEG to clean EEG without heavily relying on hand-crafted priors. Zhang et al. [8] established endto-end baselines using fully connected networks (FCNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Subsequent studies explored more advanced architectures, including the transformer-based EEGDNet [9], the embedding-separation framework DeepSeparator [10], the U-Net-based intepretable archiecture LRR-Unet [11], GANbased approaches for artifact suppression [12], [13], etc. Although these methods have shown strong denoising performances, several limitations has remained, suggesting rooms for prospective improvements. First, frequency-aware modeling of EEG artifacts is still underexplored. Different artifact sources exhibit markedly different spectral characteristics (e.g. EOG are typically concentrated in low-frequency bands [14], whereas EMG are broaderband and often more prominent in higher frequencies [15]). Explicitly modeling such band-dependent patterns can therefore provide more targeted and effective denoising. Second, a large number of existing denoisers apply relatively uniform processing over the entire signal [16], which may undercorrect severely contaminated regions or over-smooth clean segments. This motivates the need for an adaptive mechanism that can track artifact intensity over time while remaining aware of its frequency-specific behavior. Third, current deep models often emphasize either temporal sequence modeling or transformed-domain processing [16], but rarely integrate both in a manner tailored to EEG artifacts. Since EEG is inherently both temporally dynamic and spectrally structured, time-domain modeling capture temporal continuity but may overlook band-specific spectral structure, whereas frequencycentered approaches can isolate artifact-dominant bands but may weaken global temporal dependencies when bands are treated too independently. Motivated by these observations, we propose BandRouteNet, an adaptive EEG denoising network that combines frequencyaware band-wise processing with full-band temporal conditioning. The model contains two cooperative components: a Band-specific Denoiser, which operates on decomposed EEG
bands to suppress artifacts in a frequency-aware manner, and a Full-band Conditioner, which processes the original fullband signal to provide global temporal guidance and crossband contextual information. For band-wise denoising, instead of enforcing uniform denoising over all temporal regions, we further introduce an artifact routing mechanism that adaptively modulates the denoising strength according to the temporal and spectral characteristics of contamination. This design is intended to preserve cleaner neural structure while allocating stronger correction to severely corrupted band–time regions. The main contributions of this work are as followed: • First, we propose BandRouteNet, a frequency-aware and temporally adaptive denoising architecture that integrates band-specific processing with full-band contextual conditioning, enabling the exploration of both spectral structure and temporal dynamics for EEG artifact removal. • Second, we introduce an adaptive artifact routing mechanism that learns soft, time-varying denoising emphasis across latent band features, making the denoising process more selective, interpretable, and suitable for nonstationary artifact patterns. • Through extensive experiments on the EEGDenoiseNet benchmark, we show that the proposed method achieves superior denoising performance over other methods across EOG, EMG, and mixed-noise conditions under unified evaluation settings, while remaining highly parameter-efficient, supporting its potential for practical EEG denoising.
routed latent representation that adaptively controls denoising strength across time and frequency bands. The routed band features are then passed through a cross-band fusion module to model dependencies among frequency bands. The refined band features are decoded and aggregated to form a bandwise reconstruction of the clean EEG. In parallel, the full-band pathway generates an coarse-grain signal-level refinement and extracts conditional parameters from the original noisy signal to modulate the band-wise features with global contextual information. The final denoised output is obtained by adaptively fusing the outputs of these two pathways. B. Full-band Conditioner The Full-band Conditioner provides global temporal and cross-band context via adaptive feature modulation and signallevel refinement for the band-wise denoising pathway. This branch directly models the original noisy signal x ∈ RB×1×T and provides global guidance for more coherent denoising. First, the full-band noisy signal is encoded the input into a latent representation: hf = Ef (x),
hf ∈ RB×C×T ,
(1)
where Ef (·) denotes the encoder. The latent features are then processed by a temporal enhancement block to capture longrange dependencies: zf ∈ RB×C×T .
(2)
Tf (.) = Conv1D (GRU(Norm(.)))
(3)
zf = Tf (hf ), where
II. M ETHOD A. The Overall Architecture The proposed BandRouteNet is a dual-path EEG denoising framework that integrates band-wise decomposition with fullband contextual modeling. Given a noisy EEG segment X ∈ RB×1×T , where B and T denote the batch size and segment length, respectively, the network estimates a clean reconstruction Ŷ ∈ RB×1×T through two complementary components: a Band-specific Denoiser and a Full-band Conditioner. As shown at Figure 1a, the proposed BandRouteNet analyses the input EEG from two complementary views. First, the signal is decomposed into multiple frequency bands to explicitly capture band-dependent artifact characteristics. Second, the original full-band signal is modeled directly to exploit global temporal structure and inter-band context. Based on these two views, denoising is performed cooperatively: the band-specific pathway focuses on targeted artifact suppression within individual bands, while the full-band pathway provides global guidance to enhance and refine the band-wise reconstruction. The band-specific pathway follows an encode–route– reconstruct paradigm. Each decomposed band is first mapped into a latent representation, from which a band denoiser produces a candidate refined feature and an Artifact Router estimates the degree to which this refinement should be applied via a routing mechanism. Their interaction yields a
Here, zf serves as the global contextual representation of the noisy EEG. From zf , the branch produces three types of outputs. First, a decoder generates a coarse full-band refinement signal. This signal provides a signal-level correction complementary to the band-wise reconstruction: df = Df (zf ),
df ∈ RB×1×T .
(4)
Second, a temporal gating head predicts λ = σ(Hλ (zf )),
λ ∈ RB×1×T
(5)
where Hλ (.) = Conv1D (GELU(Norm(Conv1D(.))))
(6)
σ(·) denotes the sigmoid activation. The gate λ controls how much the full-band refinement temporally contribute to the final denoised output at the final fusion stage. Third, a Conv1D-based projection head outputs feature modulation parameters: [τ , ψ] = Hproj (zf ),
(7)
where τ , ψ ∈ RB×C×T . These parameters are utilized to condition the band latent features at the band-specific pathway.
(B, 1, T)
Full-band Conditioner
Decoder
Temporal Enhancement
Encoder
Time domain Noisy signal
(B, 1, T)
Latent
Gating mask:
Gating head
(B, C, T)
Proj head
(B, 1, T)
(B, 1, T)
Conv1D-Norm-GELU
GroupNorm
Band Adapter
(B, K, T)
Band Adapter Band Adapter
(B, K, C, T)
Band Adapter Gating mask:
(B, 1, T)
Scale:
(B, C, T)
Shift:
~
(B, C, T)
(B, C, T)
(B, C, T)
Band identity embedding
(B, C, T)
Decoder
Conv1D-Norm-GELU
Conv1D
(B, K, T)
Artifact Router
Band Mixer
GroupNorm
(B, K, C, T)
Encoder - Decoder
Global context
Temporal context
AdaptiveAvgPool
DW Conv1D
Conv1D
PW Conv1D
Conv1D
(B, C, T)
Conv-1D-Norm-Act
Conv-1D-Norm-Act [k=1]
Inception-1D
Inception-1D
….
Conv1D[k=3] -Norm-Act
Inception-1D
….
Conv1D
Band Adapter
(B, C, T)
Band Denoiser
(B* T, K, C)
(B, C, T)
(B, C, T)
Band-specific denoiser
(B, C, T)
FiLM
(B*K, C, T)
MHSA
ID FT
(B, C, T)
Artifact Router
Temporal Mixer
GroupNorm
FFN
Cross-band Fusion
Encoder
Concatenate
Band Adapter
(B*K, C, T)
GroupNorm Temporal Enhancement
Band Adapter
DFT
Cross-band fusion
Temporal Enhancement
GRU
(B, 1, T)
(B, C, T)
Shift:
Band Denoiser
Aggrerated Band denoised signal
(B, C, T)
Scale:
Frequency domain
Final Fusion
(B, K, C, T)
(B, C, T)
Final denoised signal
Coarse denoised signal
Conv1D
Sigmoid
(1, C, T)
Conv1D[k=5] -Norm-Act
Conv1D[k=7] -Norm-Act
Max Pooling-1D
Concatenate Conv-1D-Norm-Act [k=1]
(B, C, T)
(a) Overall framework of BandRouteNet.
(b) Detailed structures of the main component blocks.
Fig. 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 to capture global temporal dynamics and cross-band dependencies, thereby guiding and enhancing band-wise denoising.
C. Band-specific Denoiser The Band-specific Denoiser exploits the spectral heterogeneity of EEG artifacts by processing each frequency band independently, rather than forcing a single shared operation to handle all frequencies uniformly in the time domain. DFT decomposition: First, the input noisy signal X ∈ RB×1×T is decomposed into K frequency bands using an Discrete Fourier Transform (DFT) decomposition module. Specifically, for each signal x ∈ RT , the DFT is computed as x̂[f ] =
T −1 X
x[t]e−j2πf t/T ,
f = 0, 1, . . . , T − 1.
(8)
t=0
Then, the spectrum is divided into K bands by band masks {Mk }K k=1 , and each band-specific signal is reconstructed by xk = IDFT Mk ⊙ x̂ , k = 1, 2, . . . , K. (9) Thus, the band-specific representation is written as B×K×T
Xb = {x1 , x2 , . . . , xK } ∈ R
.
uk ∈ RB×C×T
(13)
Through FiLM, the global are leveraged to adapt the scale and shift of each band-specific channel over time. The conditioned features are then passed to two parallel modules: an Artifact Router and a Band Denoiser. b) The Artifact Router: This module estimates the denoising strength for each band-specific latent feature at every temporal location. Given the conditioned latent representation ek ∈ RB×C×T , the router predicts a soft routing mask: gk ∈ [0, 1]B×C×T .
(14)
(10)
(11)
Band Adapter: This module denoises the latent representation of each decomposed frequency band. Although one band adapter is applied to each band, the Artifact Router and Band Denoiser share parameters across all bands. Given the encoder features from k-th band uk , we add a learnable band identity embedding bk ∈ R1×C×1 to explicitly distinguish it when different frequency bands in the shared latent space: ũk = uk + bk
ek = FiLM(ũk , τ , ψ) = ũk ⊙ (1 + tanh(τ )) + ψ
gk = R(ek ),
The Band-specific Denoiser processes each decomposed band independently in a shared latent space. For the k-th band, the individual band input xk ∈ RB×1×T is mapped into a latent representation by a shared encoder: uk = Eb (xk ),
a) Feature-wise Linear Modulation: To inject global context information from full-band signal into bandspecific processing, we adopt Feature-wise Linear Modulation (FiLM) [17]. Given the latent representation of the k-th band, referred as uk , and the modulation parameters (τ , ψ) produced by the Full-band Conditioner, the conditioned feature is computed as
(12)
To capture the temporal dynamics and context dependence of EEG artifacts, the Artifact Router jointly models local temporal context and global channel context. The local branch employs 1D depthwise-pointwise convolutions, while the global branch uses temporal average pooling followed by convolution layers. Their outputs are fused and passed through a sigmoid activation to produce the routing mask. This design aim to yield a time-varying, channel-aware control signal that determines where and to what extent denoising should be applied. c) The Band Denoiser: This module is responsible for generating refined latent features for each decomposed band. Given the conditioned band latent ek ∈ RB×C×T , the denoiser predicts fk = B(ek ), fk ∈ RB×C×T (15)
The module comprises a pre-convolution block, a GRU-based temporal layer, and a post-convolution block within a residual framework. For the k-th band, the Band Denoiser generates a candidate refinement feature fk rather than directly replacing the input representation. This design decouples refinement generation from refinement selection, enabling the network to learn artifact-suppressed features while allowing the Routing Mechanism to adaptively determine the extent to which fk is applied, attenuated, or suppressed. d) The Routing Mechanism: The routing mechanism combines the original conditioned feature ek , the denoised proposal fk , and the routing mask gk to generate the final routed latent representation: zk = (1 − gk ) ⊙ ek + gk ⊙ fk ,
(16)
where ⊙ denotes element-wise multiplication. This formulation performs adaptive interpolation between the original feature and the denoised proposal: gk → 0 preserves the original representation, whereas gk → 1 emphasizes stronger denoising. Consequently, gk acts as a data-dependent control signal that regulates denoising strength for each band. Cross-band fusion: To model these inter-band dependencies, we employ a cross-band fusion module after band-wise routing. Given the stacked band features Z ∈ RB×K×C×T , the module refines them using two cascaded operations including a temporal mixing layer and a band mixing layer. First, temporal mixing is applied independently within each band to model intra-band temporal dynamics: Z(t) = Mtemp (Z).
(17)
For each time step, the K band features are processed by multi-head self-attention block (MHSA) [18]. Specifically, for the m-th head, Qm K⊤ m √ Vm , (19) Headm = Softmax dh and the outputs of all heads are concatenated and projected to obtain the final representation of the block. D. Final fusion After cross-band fusion, each band feature is decoded into a denoised band signal: ŷk ∈ RB×1×T
(20)
and all decoded bands are concatenated to form Ŷbands ∈ RB×K×T
Ŷ = Ŷband + λ ⊙ Df
III. E XPERIMENTAL S ETTINGS A. Datasets To evaluate the denoising performance of the proposed model, we adopt the EEGDenoiseNet benchmark [8], which is widely used in deep learning-based EEG denoising studies. The dataset consists of 4515 clean EEG signals, 3400 electrooculogram (EOG) artifact signals, and 5598 electromyogram (EMG) artifact signals. Each sample has a duration of 2 seconds with a sampling rate of 256 Hz. Let x denote the clean EEG signal and N denote the artifact signal (EOG or EMG). A contaminated EEG signal y is generated as: y = x + λN, (23) where λ controls the noise intensity. The signal-to-noise ratio (SNR) is adjusted by changing the parameter λ as following: SNR = 10 log
RMS(x) RMS(λ · n)
The aggregated band-wise denoised signal Ŷband is obtained by summing over K bands.
(24)
in which the Root Mean Squared (RMS) value is defined v u N u1 X RMS(g) = t g2 (25) N i=1 i
From the samples belong to each types of signal, three types of dataset are constructed: a) EOG Dataset: The EOG-contaminated dataset is created from 3400 EEG segments and 3400 EOG segments. At specified signal-to-noise ratio (SNR) levels, pure EEG signals were linearly combined with ocular artifacts to generate EOGcontaminated signals. b) EMG Dataset: In the EMG setting, EEG signals are randomly reused to match the number of EMG segments, resulting in 5598 samples from the EOG-contaminated EEG signals. Similarly, using specified SNR levels, pure EEG signals were linearly combined with EMG artifacts to generate EMG artifact-contaminated signals. c) Mixed EOG/EMG Dataset: For the mixed-artifact setting, EEG, EOG, and EMG signals are all aligned to a common size equal to the maximum number of artifact samples. For each SNR level, both EOG/EMG are scaled independently to that level and linearly combined: y = x + λ(NEOG + NEMG )
(21)
(22)
where the condition parameter λ control the contribution of globally-denoised signal to the final denoised signal.
as
Next, in the band mixing stage, features at each time step interact across the K bands through multi-head self-attention (MHSA) block, followed by a position-wise feed-forward network: Z′ = Mband (Z(t) ). (18)
ŷk = Db (z′k ),
The final denoised signal is the fusion that combines finegrained band-wise denoising and coarse-grained global denoised signal from the Full-band conditioner as:
(26)
Following the dataset construction protocol described in the official benchmark study [8], each EEG-artifact pair is augmented across 10 signal-to-noise ratio (SNR) levels ranging
TABLE I P ERFORMANCE C OMPARISON OF AVERAGE P ERFORMANCES ACROSS A LL SNR L EVELS . T HE SMALLER RMSSET , RMSSES , AND THE L ARGER CC, SNRIMP THE B ETTER D ENOISING E FFECT. B EST R ESULTS ARE IN B OLD ; S ECOND - BEST O NES ARE U NDERLINED EOG dataset
EMG dataset
Mixed EOG/EMG dataset
Method
Parameter (M) RRMSEt
RRMSEs
CC
SNRimp
RRMSEt
RRMSEs
CC
SNRimp
RRMSEt
RRMSEs
CC
FCNN [8]
1.051
0.5570
0.5865
0.8111
10.8011
0.6176
0.6687
0.7849
9.6891
0.6887
0.7362
0.7123
11.7388
Simple-CNN [8]
16.82
0.4437
0.4432
0.8847
12.6351
0.7215
0.7263
0.7255
8.4879
0.7787
0.7258
0.6763
10.7515 10.0200
SNRimp
1D-ResCNN [19]
8.46
0.4184
0.4046
0.8966
13.1761
0.7510
0.7093
0.7076
8.2550
0.8674
0.9065
0.6625
RNN-LSTM [8]
0.788
0.6903
0.7231
0.7176
8.6099
0.6758
0.7215
0.7384
8.8269
0.7522
0.8123
0.6644
10.8124
EEGDnet [9]
0.895
0.4406
0.4125
0.8907
12.4712
0.6554
0.6196
0.7456
9.0505
0.6897
0.6210
0.7125
11.5704
Deep Separator [10]
0.032
0.4944
0.5484
0.8658
11.3920
0.6982
0.6899
0.7396
8.6998
0.7593
0.7174
0.6814
10.7386
LRR-UNet [11]
3.19
0.4322
0.4311
0.8967
12.7577
0.6415
0.5827
0.7564
9.3746
0.7144
0.6328
0.7068
11.3037
BandRouteNet (Ours)
0.20
0.3831
0.3797
0.9156
13.9819
0.5962
0.5276
0.7802
10.0041
0.6605
0.5998
0.7359
11.9310
TABLE II A BLATION S TUDY R ESULTS ON EOG DATASET Experiments Full model W/o fullband conditioner W/o artifact routing (g = 1) W/o cross-band fusion W/o band identity embedding
RRMSEt 0.3831 0.3907 0.4091 0.3956 0.3867
RRMSEs 0.3797 0.3979 0.4013 0.4009 0.3998
CC 0.9156 0.9106 0.9112 0.9139 0.9130
SNRimp 13.9819 13.2324 13.7610 13.8310 13.8392
from −7 to 2 dB to enhance training diversity. The resulting datasets are then divided into training, validation, and test sets with a ratio of 8:1:1. Furthermore, both clean and noisy signals are standardized using the standard deviation of the corresponding noisy signal, as defined below: x̂ =
x , σy
ŷ =
y σy
(27)
B. Evaluation Metrics To quantitatively assess the denoising performance of the proposed model, we adopt the standard evaluation metrics, including the Temporal Root Mean Squared Error (RRMSEt ), Spectral Root Mean Squared Error (RRMSEs ), and Correlation Coefficient (CC). In addition, we also report the Signal-tonoise Ratio Improvement (SNRimp ) to measure the gain in signal quality before and after denoising. a) Temporal Relative Root Mean Squared Error: RRMSEt evaluates the waveform-level reconstruction error between the denoised signal ŷ and the clean reference signal y: RMS(ŷ − y) RRMSEt = (28) RMS(y) b) Spectral Relative Root Mean Squared Error: RRMSEs measures the discrepancy between the spectra of the denoised and clean signals: RRMSEs =
RMS(S(ŷ) − S(y)) RMS(S(y))
(29)
where S(·) denotes the Power Spectral Density (PSD) of the signal.
c) Correlation Coefficient: The correlation coefficient (CC) measures the linear similarity between the denoised signal and the clean target: PN ¯ i=1 (yi − ȳ) ŷi − ŷ q (30) CC = qP N 2 PN ¯ 2 ŷ − ŷ (y − ȳ) i i i=1 i=1 where ȳ and ŷ¯ denote the mean values of y and ŷ, respectively. d) SNR improvement: To further evaluate denoising effectiveness, we compute the SNR improvement (SNR imp), defined as the difference between the output SNR and the input SNR: SNRimp = SNRout − SNRin
(31)
Given the clean signal x, noisy observation y, and denoised output ŷ, the input and output SNR values are computed as P (x) SNRin = 10 log10 , P (y − x) (32) P (x) . SNRout = 10 log10 P (ŷ − x) where P (·) denotes the average signal power. C. Implementation Details The proposed model was implemented in PyTorch and trained on an NVIDIA Tesla T4 GPU. We used the AdamW optimizer [20] with an initial learning rate of 1 × 10−3 and a weight decay of 1 × 10−4 . The model was trained for 15 epochs using Mean Squared Error (MSE) as the loss function. For model configuration, the input noisy EEG signal was decomposed into K = 6 frequency bands: Delta (δ, 0—4 Hz), Theta (θ, 4—8 Hz), Alpha (α, 8—12 Hz), Beta (β, 13—30 Hz), Gamma (γ, 30—80 Hz), and a higher band ϵ (80—128 Hz). The hidden feature dimension was set to C = 64. Both the Band-specific Denoiser and the Full-band Conditioner adopt an Inception-based encoder—decoder architecture to capture multi-scale features, as shown at the Figure 1bb. Each module consists of two encoder stages and two decoder stages, with each stage incorporating two Inception1D blocks.
Deep Separator LRR-Unet BandRouteNet
Fig. 4. Denoising performance comparison on the mixed-noise segment
EEGDnet
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
0.7
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
0.6 0.5 0.4 0.3
0.5 0.4
6
4 2 SNR Level (dB)
0
0.3
2
A. Qualitative Evaluation in the Time and Frequency Domains We first conduct visual analysis to compare denoising performances of all methods under EOG, EMG, and mixed-noise conditions using reconstructed waveforms and PSD curves, as shown in Fig. 2, Fig. 3, and Fig. 4. Across all scenarios, BandRouteNet produces reconstructions that more closely follow the clean target (depicted in green) while effectively suppressing artifacts without introducing residual noise or over-smoothing. Its PSD curves also better match the clean spectra (depicted in green), indicating stronger preservation of EEG spectral characteristics. These results suggest that BandRouteNet achieves a favorable balance between artifact removal and intrinsic EEG signal preservation. B. Quantitative Comparison Across Evaluation Metrics We conduct quantitative experiments to compare the denoising performance of the proposed BandRouteNet against all competing methods. Table I reports the average results across all SNR levels on the three benchmark settings, namely
0.95
20.0
0.90
17.5 SNR Improvement (dB)
CC
To assess the effectiveness of the proposed BandRouteNet, we compare it against several neural network-based EEG denoising methods, including the benchmark baselines FCNN [8], SimpleCNN [8], RNN-LSTM [8], and 1DResCNN [8], as well as recent state-of-the-art architectures such as EEGDNet [9], Deep Separator [10], and LRRUnet [11]. To ensure a fair comparison, all competing methods are reproduced and evaluated under a unified experimental protocol, including the same training, validation, and test splits, identical evaluation and visualization settings.
0.80 0.75 0.70
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
0.65 6
4 2 SNR Level (dB)
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
0
(c) Correlation coefficient
6
4 2 SNR Level (dB)
0
2
(b) RMSSE spectral
0.85
IV. R ESULTS AND D ISCUSSION
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
0.6
(a) RMSSE temporal Fig. 3. Denoising performance comparison on the EMG-contaminated segment
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
0.7 RMSSE_s
RMSSE_t
Deep Separator LRR-Unet BandRouteNet
EEGDnet
FCNN Simple CNN Res-1D-CNN RNN-LSTM
EEGDnet Deep Separator LRR-Unet BandRouteNet
FCNN Simple CNN Res-1D-CNN RNN-LSTM Res-1D-CNN RNN-LSTM
Simple CNN
FCNN
Fig. 2. Denoising performance comparison on the EOG-contaminated segment
2
15.0 12.5 10.0 7.5 5.0 2.5 0.0
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
6
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
4 2 SNR Level (dB)
0
2
(d) SNR improvement
Fig. 5. Performance Metrics on different SNR Levels (EOG Dataset)
the EOG, EMG, and mixed EOG/EMG datasets. The evaluation follows four standard metrics, including RMSSEt , RMSSEs , CC, and SNRimp . As shown in Table I, the proposed model achieves the best overall performance across all three datasets. On the EOG dataset, BandRouteNet attains the lowest RMSSEt (0.3831) and RMSSEs (0.3797), as well as the highest CC (0.9156) and SNRimp (13.9819), outperforming all competing methods. A similar trend is observed on the EMG dataset, where the proposed model again yields the best RMSSEt (0.5962), RMSSEs (0.5276), and SNR improvement (10.0041), while remaining highly competitive in CC (0.7802). On the mixed EOG/EMG dataset, which is the most challenging setting due to the coexistence of multiple artifact types, the proposed system continues to deliver the strongest overall performance, achieving the best RMSSEt (0.6605), RMSSEs (0.5998), CC (0.7359), and SNRimp (11.9310). These results demonstrate that the proposed architecture generalizes well across different contamination types and provides a consistently better trade-off between waveform fidelity, spectral preservation, correlation with the clean signal, and noise suppression. We further analyze the performance across different SNR
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
1.2
1.6
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
1.4
1.0
RMSSE_s
0.8
0.8
0.6
0.6 0.4
0.4 6
4 2 SNR Level (dB)
0
2
6
(a) RMSSE temporal 16 12
SNR Improvement (dB)
0.8 CC
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
14
0.7 0.6 0.5
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
6
4 2 SNR Level (dB)
0
2
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
10 8 6 4
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
0
4 2 SNR Level (dB)
(b) RMSSE spectral
0.9
0.4
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
1.2
1.0 RMSSE_t
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
2
2
6
(c) Correlation coefficient
4 2 SNR Level (dB)
0
2
(d) SNR improvement
Fig. 6. Performance Metrics on different SNR Levels (EMG Dataset)
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
1.4
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
1.75
1.0 0.8
1.25 1.00 0.75
0.6
0.50 6
4 2 SNR Level (dB)
0
2
6
(a) RMSSE temporal 18 SNR Improvement (dB)
0.7 CC
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
16
0.8
0.6 0.5 FCNN SimpleCNN 1D-ResCNN RNN-LSTM
0.4 6
4 2 SNR Level (dB)
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
0
(c) Correlation coefficient
4 2 SNR Level (dB)
0
2
(b) RMSSE spectral
0.9
0.3
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
1.50 RMSSE_s
RMSSE_t
1.2
0.4
FCNN SimpleCNN 1D-ResCNN RNN-LSTM
2.00
2
EEG-DNet DeepSeparator LRR-UNet BandRouteNet (Ours)
stable and favorable performance, further demonstrating its generalization capability across different noise intensities. In addition to the strong denoising performance, BandRouteNet is also highly lightweight. The proposed model contains only 0.20 million trainable parameters, which is far smaller than several comparing methods, yet it still delivers clearly superior performance. Overall, these quantitative results show that BandRouteNet is not only more accurate than other methods, but also more parameter-efficient, showing potential to be a practical solution for high-performance EEG denoising under resource-constrained settings. C. Ablation Analysis of the Proposed Network Components To assess the contribution of each module to the model performance, we perform an ablation study on the EOG dataset by removing one component at a time from the full model, including the full-band conditioner, artifact routing gate, cross-band fusion module, and band identity embedding. The results reported in Table II show that each proposed component contributes to the final denoising quality. While the full model achieves the best performance on all metrics, with RMSSEt of 0.3831, RMSSEs of 0.3797, CC of 0.9156, and an SNRimp of 13.9819, all ablated variants show degraded performance. Notably, removing the artifact routing gate leads to the largest drop in performance (increases of 0.026 and 0.0216 in RMSSEt and RMSSEs , respectively, indicating that adaptive temporal routing is a key factor in the proposed architecture. Removing the full-band conditioner also causes a clear reduction in performance (with RMSSEt and RMSSEs increasing by 0.0076 and 0.0182, respectively), which verifies the importance of incorporating global temporal and cross-band contextual information in addition to bandwise processing. Other components such as cross-band fusion and band identity embedding positively contribute to the better reconstruction of artifact-free EEG signals.
14 12
D. Visualization of the Artifact Routing Mechanism
10 8 6 4 6
4 2 SNR Level (dB)
0
2
(d) SNR improvement
Fig. 7. Performance Metrics on different SNR Levels (Mixed EOG/EEG Dataset)
levels, as shown in Fig. 5, Fig. 6, and Fig. 7. In general, all methods improve when the input SNR increases. Specifically, BandRouteNet maintains superior or highly competitive performance throughout the full SNR range. In particular, the proposed model consistently achieves lower temporal and spectral reconstruction errors, while preserving higher or competitive correlation and SNR improvement over most SNR levels. This advantage is especially evident in the lowSNR regime, where the denoising task is most challenging. As the SNR increases, the proposed model continues to exhibit
To analyze the proposed routing mechanism interpretably, we visualize the Artifact Router outputs under EOG, EMG, and mixed EOG/EMG noise conditions in Fig. 8. In each subfigure, the upper panel shows the denoising result, while the lower panel presents the corresponding router activation map. Larger heatmap values indicate stronger denoising emphasis on the corresponding band-specific features. The activations vary adaptively across both time and frequency. For EOG noise, strong responses mainly appear in low-frequency bands, especially Delta (δ, 0–4 Hz) and Theta (θ, 4–8 Hz), during intervals with prominent ocular artifacts. For EMG noise, the responses shift toward higher-frequency bands, consistent with the broadband, high-frequency nature of muscle artifacts. Under mixed noise, the routing map combines both patterns, showing that the model can respond to coexisting low- and high-frequency contamination. The activations are also temporally localized and smooth, suggesting that the router increases denoising mainly in corrupted regions while preserving cleaner signal segments.
(a) EOG-contaminated EEG signal
(b) EMG-contaminated EEG signal
(c) EOG/EMG-contaminated EEG signal
Fig. 8. Artifact routing visualization results under three noise conditions. In each sub-figure, the upper panels present the denoising results, while the lower panel shows the outputs of the Artifact Router module as a heatmap over time and frequency bands.
V. C ONCLUSION This paper proposed BandRouteNet, an adaptive neural network for EEG denoising. By combining band-specific denoising with routing-based adaptive refinement, complemented by full-band contextual conditioning, the proposed model effectively suppresses diverse artifacts and preserve the intrinsic structure of EEG signals. Extensive experiments under various noise conditions show that BandRouteNet achieves superior performance across various standard evaluation metrics, while remaining highly parameter-efficient. R EFERENCES [1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
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. 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. 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. S. Halder et al., “Online artifact removal for braincomputer interfaces using support vector machines and blind source separation,” Computational intelligence and neuroscience, vol. 2007, no. 1, p. 082 069, 2007. 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. C. Kaur, P. Singh, and S. Sahni, “Eeg artifact removal system for depression using a hybrid denoising approach,” Basic and Clinical Neuroscience, vol. 12, no. 4, p. 465, 2021. 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. H. Zhang et al., Eegdenoisenet: A benchmark dataset for end-to-end deep learning solutions of eeg denoising, 2021. arXiv: 2009.11662 [eess.SP].
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18] [19]
[20]
X. Pu et al., “Eegdnet: Fusing non-local and local self-similarity for eeg signal denoising with transformer,” Computers in Biology and Medicine, vol. 151, p. 106 248, 2022, ISSN: 0010-4825. 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. 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. E. Brophy et al., “Denoising eeg signals for realworld bci applications using gans,” Frontiers in Neuroergonomics, vol. Volume 2 - 2021, 2022, ISSN: 26736195. DOI: 10.3389/fnrgo.2021.805573 H. Wang et al., “Eeg signal denoising using pix2pix gan: Enhancing neurological data analysis,” arXiv preprint arXiv:2411.13288, 2024. 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. I. I. Goncharova et al., “Emg contamination of eeg: Spectral and topographical characteristics,” Clinical neurophysiology, vol. 114, no. 9, pp. 1580–1593, 2003. V. A. Raj et al., “A comprehensive review of deep learning models for denoising eeg signals: Challenges, advances, and future directions,” Discover Applied Sciences, vol. 7, no. 11, p. 1268, 2025. E. Perez et al., “Film: Visual reasoning with a general conditioning layer,” in Proceedings of the AAAI conference on artificial intelligence, vol. 32, 2018. A. Vaswani et al., Attention is all you need, 2023. arXiv: 1706.03762 [cs.CL]. 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. I. Loshchilov and F. Hutter, Decoupled weight decay regularization, 2019. arXiv: 1711.05101 [cs.LG].