EEGDenoiseNet: EEG Denoising Benchmark

• EEGDenoiseNet is a standardized benchmark for EEG denoising that provides curated clean EEG segments alongside ocular and myogenic artifact pools for rigorous evaluation.
• It employs a semi-synthetic contamination protocol, where clean and artifact segments are linearly mixed to create controlled noisy inputs with defined signal-to-noise ratios.
• The framework has been pivotal in comparing diverse deep learning models—from FCNNs and CNNs to LSTMs and GANs—across single-channel denoising tasks.

Searching arXiv for EEGDenoiseNet and closely related benchmark papers to ground the article in current literature. EEGDenoiseNet is a benchmark dataset and evaluation framework for supervised, end-to-end EEG denoising, introduced to provide standardized clean EEG targets, isolated artifact segments, and common benchmarks for deep-learning-based artifact removal (Zhang et al., 2020). It was created to address the absence of an open dataset explicitly structured for denoising research, where contaminated EEG can be synthesized with known ground-truth clean EEG, allowing fair comparison across models and controlled study of denoising under ocular and myogenic contamination (Zhang et al., 2020). In its original form, EEGDenoiseNet contains curated single-channel 2-second segments of clean EEG, ocular artifact, and muscular artifact, and has subsequently become a reference benchmark for studies on ocular artifact removal, EMG suppression, frequency-conditioned denoising, Transformer- and RetNet-based sequence models, GAN-based denoising, and capacity-controlled benchmark analysis (Zhang et al., 2020, Gabardi et al., 2023, Yi et al., 2021, Wang et al., 2024, Wang et al., 2024, Bindra et al., 7 Jun 2026).

1. Dataset definition and construction

EEGDenoiseNet was introduced as a publicly available benchmark dataset specifically for deep-learning EEG denoising, rather than for general EEG analysis or BCI tasks (Zhang et al., 2020). Its defining feature is that it separates the data needed for supervised denoising into three curated pools: 4514 clean EEG segments, 3400 ocular artifact segments, and 5598 myogenic artifact segments (Zhang et al., 2020). Each segment is 2 seconds long, and the dataset is distributed in both MATLAB .mat and Python NumPy .npy formats (Zhang et al., 2020).

The clean EEG pool was derived from the motor imagery dataset of Cho et al. using 52 participants, 64-channel EEG, and a 512 Hz sampling rate; those recordings were processed with 1–80 Hz band-pass filtering, a notch filter at the local powerline frequency, resampling to 256 Hz, and artifact attenuation using ICLabel, followed by segmentation into 2 s one-dimensional segments, standardization, and expert visual inspection (Zhang et al., 2020). The ocular artifact pool was extracted from open datasets containing horizontal and vertical EOG channels, with preprocessing consisting of 0.3–10 Hz band-pass filtering, resampling to 256 Hz, segmentation into 2 s one-dimensional segments, standardization, and expert inspection (Zhang et al., 2020). The myogenic artifact pool was derived from facial EMG recordings, processed with 1–120 Hz band-pass filtering, a notch filter at powerline frequency, resampling to 512 Hz, segmentation into 2 s segments, standardization, and expert visual inspection (Zhang et al., 2020).

The dataset is intentionally single-channel, which the original paper presents as a design choice to improve universality across EEG systems rather than tie the benchmark to a fixed multichannel montage (Zhang et al., 2020). The 2-second segment length was chosen because it is long enough to preserve temporal and spectral characteristics of EEG, EOG, and EMG, yet short enough that artifact-free clean EEG segments can realistically be curated (Zhang et al., 2020).

2. Semi-synthetic contamination protocol

EEGDenoiseNet is fundamentally a semi-synthetic supervised denoising benchmark. The released dataset does not store noisy-clean pairs directly; instead, it provides curated clean EEG and artifact segments from which contaminated EEG can be synthesized by linear mixing (Zhang et al., 2020). The canonical contamination model is

$y = x + \lambda \cdot n$

where $y$ is the contaminated EEG, $x$ is the clean EEG target, $n$ is an artifact segment, and $\lambda$ controls contamination strength and thus the signal-to-noise ratio (Zhang et al., 2020). The signal-to-noise ratio is defined through the RMS ratio between clean EEG and scaled artifact, with the RMS itself defined as

$RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$

in the benchmark paper (Zhang et al., 2020).

This protocol creates exact supervised pairs: the model input is contaminated EEG $y$, and the training target is the clean EEG $x$ (Zhang et al., 2020). For ocular artifact experiments in the original benchmark, 3400 EEG segments and 3400 ocular artifact segments are split into 80% training, 10% validation, and 10% test, then expanded by random linear mixing across 10 SNR levels from $-7$ dB to $2$ dB (Zhang et al., 2020). For myogenic artifact experiments, the EEG pool is upsampled to 512 Hz to match EMG, and some EEG segments are randomly reused because there are fewer clean EEG than EMG segments (Zhang et al., 2020).

A second normalization stage is applied at model-training time: both clean and contaminated signals are normalized by the standard deviation of the contaminated signal, in the intended form

$y$0

with the standard deviation retained for rescaling the denoised output back to the original amplitude domain (Zhang et al., 2020). This contaminated-signal-based normalization was reused in later EEGDenoiseNet studies, including the EMG-focused “Novel CNN” work and the RetNet-based EEGDiR model (Zhang et al., 2020, Wang et al., 2024).

A central methodological consequence is that EEGDenoiseNet evaluates denoisers under a controlled additive model with known ground truth. This makes the benchmark well suited to supervised reconstruction studies, but also constrains conclusions to the semi-synthetic setting. Later work repeatedly notes that success on EEGDenoiseNet does not by itself establish generalization to real contaminated EEG, where artifact coupling may be nonlinear, electrode-dependent, or mixed across multiple artifact classes (Ozdemir et al., 2022, Wang et al., 2024, Bindra et al., 7 Jun 2026).

3. Original benchmark models and evaluation protocol

The original EEGDenoiseNet paper provides not only the dataset but also a benchmark protocol with four deep-learning model families: a fully connected neural network (FCNN), a simple CNN, a complex CNN, and an LSTM-based RNN (Zhang et al., 2020). All are trained end-to-end with mean squared error under the mapping

$y$1

from normalized contaminated input to normalized clean target (Zhang et al., 2020). Training used Adam with $y$2, $y$3, and $y$4, implemented in Python 3.7 and TensorFlow 2.2 on two NVIDIA Tesla V100 GPUs (Zhang et al., 2020).

The FCNN has 4 hidden layers with ReLU and dropout, with layer width equal to input length: 512 neurons for ocular denoising and 1024 neurons for myogenic denoising (Zhang et al., 2020). The simple CNN contains 4 one-dimensional convolution layers with kernel $y$5, stride 1, 64 feature maps, batch normalization, ReLU, and a flatten-plus-dense reconstruction head (Zhang et al., 2020). The complex CNN is a 1D-ResCNN with parallel residual branches using kernels $y$6, $y$7, and $y$8 (Zhang et al., 2020). The RNN benchmark is an LSTM whose outputs are passed through a three-layer fully connected network with ReLU and dropout (Zhang et al., 2020).

Evaluation is based on convergence curves, temporal-domain relative RMSE, spectral-domain relative RMSE, and correlation coefficient. The original benchmark defines temporal and spectral error as

$y$9

and

$x$0

with correlation measured by the covariance-normalized coefficient $x$1 (Zhang et al., 2020). The benchmark also includes two traditional single-channel baselines, empirical mode decomposition and filtering, and reports repeated randomized experiments across 10 independently generated datasets for statistical robustness (Zhang et al., 2020).

The benchmark’s headline findings are artifact-dependent. For ocular artifact removal, the RNN is reported to have the lowest RRMSE and the highest CC among the four deep models; for myogenic artifact removal, the complex CNN performs best overall (Zhang et al., 2020). The paper further reports that deep-learning methods outperform EMD and filtering for both ocular and myogenic denoising, with $x$2 in ANOVA-based comparisons for temporal RRMSE, spectral RRMSE, and CC (Zhang et al., 2020).

4. Benchmark behavior, artifact asymmetry, and methodological scope

A persistent feature of EEGDenoiseNet is the asymmetry between ocular and myogenic denoising. The original benchmark paper explicitly associates ocular artifacts with lower-frequency contamination and myogenic artifacts with broader, higher-frequency spectral structure (Zhang et al., 2020). Later studies sharpen this interpretation. The frequency-based deep-learning model of 2023 characterizes ocular activity as low-frequency, high-amplitude activity around 0.5–3 Hz, whereas muscle activity overlaps more broadly with EEG and is usually $x$3 Hz, making EMG denoising harder because it risks removing genuine neural content (Gabardi et al., 2023). This suggests that EEGDenoiseNet does not represent a single homogeneous denoising task, but rather at least two benchmark regimes with different spectral separability and error profiles.

This asymmetry is visible across later models. The frequency-based dynamic-filtering approach reported in 2023 achieves its strongest result on ocular artifact removal, where it reports $x$4, $x$5, and $x$6, while remaining only competitive on muscle artifact removal with $x$7, $x$8, and $x$9 (Gabardi et al., 2023). EEGDiR likewise reports larger margins over prior methods on EOG and SS2016 EOG than on EMG, with average EEGDenoiseNet EOG results of $n$0 for $n$1, $n$2, and $n$3, versus $n$4 on EMG (Wang et al., 2024).

The benchmark’s scope is also limited by construction. EEGDenoiseNet is built around single-artifact contamination, conventionally evaluated separately for EOG and EMG. Even when later work introduces “multi-artifact” denoisers, the term often refers to one shared model trained across separate EOG-corrupted and EMG-corrupted samples, not simultaneous contamination by both artifacts in the same input (Gabardi et al., 2023). Only more recent work such as BandRouteNet explicitly constructs a mixed EOG/EMG dataset on top of EEGDenoiseNet and evaluates that case under a unified protocol (Lam, 27 Apr 2026). The original benchmark therefore standardizes single-channel, single-artifact supervised denoising more than full real-world artifact ecology.

5. Subsequent model families built on EEGDenoiseNet

EEGDenoiseNet quickly became the standard evaluation setting for a broad range of neural denoisers. One important line of work sought stronger CNN baselines for the harder EMG task. The 2020 “Novel CNN” paper used 4514 EEG and 5598 EMG epochs from EEGDenoiseNet, expanded EEG by reuse to 5598 pairs, split them into 4478/560/560 train/validation/test pairs, and trained a 7-block 1D CNN with progressively increasing channels 32, 64, 128, 256, 512, 1024, 2048 and average pooling after the first six blocks (Zhang et al., 2020). On EEGDenoiseNet EMG denoising, it reported $n$5, $n$6, and $n$7, outperforming the benchmark FCNN and RNN on average, while also arguing that the new architecture largely eliminated the severe overfitting observed in earlier CNN baselines (Zhang et al., 2020).

A second line adapted sequence models. EEGDnet reshapes a 1-D EEG epoch into a $n$8 matrix, with baseline configuration $n$9 for 512-sample ocular experiments, and applies a Transformer encoder to jointly model what the paper terms non-local and local self-similarity (Yi et al., 2021). Under the EEGDenoiseNet protocol with 4514 EEG, 3400 ocular, and 5598 muscle epochs, contaminated by $\lambda$0 at uniformly distributed SNR from $\lambda$1 dB to $\lambda$2 dB, EEGDnet reports for ocular artifact removal $\lambda$3, $\lambda$4, and $\lambda$5, and for muscle artifacts $\lambda$6, $\lambda$7, and $\lambda$8 respectively (Yi et al., 2021).

EEGDiR then replaced Transformer attention with a Retentive Network-derived mechanism, using signal embedding that patchifies a 512-sample segment with patch size 16, yielding a $\lambda$9 hidden representation and 4 DiR blocks with 8 heads (Wang et al., 2024). It retains the canonical contamination model $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$0 and EEGDenoiseNet’s $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$1 dB to $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$2 dB range, and reports improved average scores over SCNN, 1D-ResCNN, LSTM, and EEGDnet on EOG, EMG, and SS2016 EOG (Wang et al., 2024).

Another line explored explicit frequency conditioning. The 2023 paper “A multi-artifact EEG denoising by frequency-based deep learning” treats EEGDenoiseNet as a controlled setting where the PSD of the pure noise $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$3 and the PSD of the noisy signal $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$4 can be used as side information to generate sample-specific denoising kernels (Gabardi et al., 2023). The network’s kernel evaluator contains two symmetric branches of 1D convolutional layers with tanh activations for the real and imaginary parts of a filter in the frequency domain, followed by IRFFT, a kernel-size-1 1D convolution, and final tanh, with total trainable parameter count 224,192 (Gabardi et al., 2023). Its distinctiveness is not merely performance but the way it exploits EEGDenoiseNet’s synthetic construction to assume access to contaminant PSD.

Generative models have also been evaluated on EEGDenoiseNet. A 2024 study recasts single-channel EMG denoising as paired image-to-image translation with pix2pixGAN, converting each 1024-sample contaminated waveform into a 32×32 image, training a conditional GAN, and reshaping the denoised image back to a time series (Wang et al., 2024). Its average results over all SNR levels show CC 0.8645, RRMSE$RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$5 0.3596, and RRMSE$RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$6 0.5797, indicating particularly strong spectral recovery and correlation, though not the best average temporal RRMSE by its own table (Wang et al., 2024). A more compact efficiency-oriented line is represented by TADA, a Targeted Adversarial Denoising Autoencoder with less than 400,000 trainable parameters, SNR-aware targeting, and covariance-guided logistic rescaling, evaluated on EEGDenoiseNet EMG removal from 67 subjects (Choi et al., 9 Jan 2025).

The benchmark has also supported extensions beyond the original dataset. One ocular-artifact study combined EEGDenoiseNet with DEAP, yielding 6164 clean EEG segments, 5000 EOG segments, and a derived 150,000 augmented sample corpus, and used BiLSTM models with raw, STFT, CWT, and WSST inputs, finding WSST-Net best with average MSE 0.3066 (Ozdemir et al., 2022). Although this differs from pure EEGDenoiseNet benchmarking, it shows the benchmark’s role as a base resource for constructing larger semi-synthetic denoising datasets.

6. Benchmark saturation, reproducibility issues, and current interpretation

Later analyses have shifted the interpretation of EEGDenoiseNet from merely a shared benchmark to an object of methodological scrutiny. The 2026 capacity-controlled study argues that the benchmark is already close to saturated with very small models (Bindra et al., 7 Jun 2026). Holding architecture family, loss, split, optimizer, and training recipe fixed, and varying only channel width in a depthwise-separable convolutional U-Net from 1.05K to 40.26K parameters, the study reports that reconstruction performance on EEGDenoiseNet saturates by roughly 3–6.5K parameters, with post-elbow gains of at most 0.015 correlation coefficient per $RMS(g) = \sqrt{\frac1{N}\sum_{i=1}^N g_i^2}$7-parameter unit (Bindra et al., 7 Jun 2026). A retrained 8.46M-parameter EEGDN CNN matches the 40.26K compact model on EOG, a roughly 200× parameter gap without reconstruction advantage under the same controlled pipeline (Bindra et al., 7 Jun 2026).

The same paper argues that strong EEGDenoiseNet reconstruction metrics do not reliably predict downstream BCI utility. Using denoisers trained under EEGDenoiseNet-style supervision and testing them on BCI Competition IV-2a motor imagery classification, it reports that reconstruction-optimized denoising can significantly degrade CSP+LDA accuracy across all nine subjects and multiple artifact conditions, including naturally recorded trials (Bindra et al., 7 Jun 2026). This does not invalidate EEGDenoiseNet as a reconstruction benchmark; rather, it narrows what the benchmark can be said to measure. A plausible implication is that EEGDenoiseNet is valuable for standardized supervised signal restoration, but insufficient as a sole proxy for preservation of task-relevant neural information.

Reproducibility is another persistent issue in EEGDenoiseNet-based literature. Many later papers note missing details in architectural specifications, data split procedures, exact PSD implementations, or subject-wise partitioning (Gabardi et al., 2023, Wang et al., 2024, Choi et al., 9 Jan 2025, Lam, 27 Apr 2026). In some studies, printed equations are malformed, especially normalization formulas, loss expressions, and correlation metrics (Gabardi et al., 2023, Ozdemir et al., 2022, Wang et al., 2024). BandRouteNet, despite strong reported results under a unified protocol and only 0.20M parameters, leaves unspecified several implementation details such as batch size, learning-rate schedule, Inception1D branch structure, GRU dimensions, and attention head counts (Lam, 27 Apr 2026). These issues have made “same-pipeline” retraining studies especially important for interpreting benchmark progress.

Taken together, EEGDenoiseNet is best understood as a standardized but constrained benchmark for single-channel, semi-synthetic, supervised EEG denoising, built around clean-reference reconstruction under controlled EOG and EMG contamination (Zhang et al., 2020). Its enduring significance lies in providing a common substrate on which denoising architectures can be compared, training regimes can be stress-tested, and broader questions about capacity, artifact-specific difficulty, and metric validity can be studied (Zhang et al., 2020, Bindra et al., 7 Jun 2026). Its limitations—single-artifact bias, additive mixing assumptions, lack of real noisy-clean pairs, and incomplete alignment with downstream utility—have become part of the benchmark’s identity rather than merely external caveats (Ozdemir et al., 2022, Wang et al., 2024, Bindra et al., 7 Jun 2026).