PaperlayerView exact revision on arXiv

Source-linked AI summary

EEGdenoiseNet: A benchmark dataset for end-to-end deep learning solutions of EEG denoising

Haoming Zhang, Mingqi Zhao, Chen Wei, Dante Mantini, Zherui Li, Quanying Liu

arXiv:2009.11662v4eess.SP

TL;DR

EEG denoising research lacks well-structured datasets with clean ground truth and standardized benchmarks. The paper introduces EEGdenoiseNet, a public dataset with controlled EEG-artifact synthesis and four end-to-end deep-learning baselines, finding that deep-learning methods outperform traditional methods for ocular and myogenic artifact removal and show potential even under high noise. The dataset is limited in data diversity and focuses on two-second segments rather than continuous EEG.

  • Problem

    The lack of ground-truth clean EEG data and specific benchmarks limits development and comparison of supervised deep-learning methods for EEG denoising.

  • Method

    EEGdenoiseNet provides clean EEG, ocular-artifact, and myogenic-artifact segments for controlled contamination synthesis, and evaluates four supervised end-to-end deep-learning networks.

  • Results

    Deep-learning methods outperform two traditional methods for ocular and myogenic artifact removal across RRMSEtemporal, RRMSEspectral, and CC, with p < 0.001 for each metric.

  • Takeaways & Limitations

    EEGdenoiseNet offers a public benchmark dataset and tools, while the results suggest deep learning has potential for removing ocular and myogenic artifacts even at high noise levels.

  • Takeaways & Limitations

    The dataset has limited EEG and artifact diversity, and the study focuses only on denoising two-second EEG segments rather than continuous EEG.

Abstract

from arXiv · show

Deep learning networks are increasingly attracting attention in various fields, including electroencephalography (EEG) signal processing. These models provided comparable performance with that of traditional techniques. At present, however, lacks of well-structured and standardized datasets with specific benchmark limit the development of deep learning solutions for EEG denoising. Here, we present EEGdenoiseNet, a benchmark EEG dataset that is suited for training and testing deep learning-based denoising models, as well as for performance comparisons across models. EEGdenoiseNet contains 4514 clean EEG segments, 3400 ocular artifact segments and 5598 muscular artifact segments, allowing users to synthesize contaminated EEG segments with the ground-truth clean EEG. We used EEGdenoiseNet to evaluate denoising performance of four classical networks (a fully-connected network, a simple and a complex convolution network, and a recurrent neural network). Our analysis suggested that deep learning methods have great potential for EEG denoising even under high noise contamination. Through EEGdenoiseNet, we hope to accelerate the development of the emerging field of deep learning-based EEG denoising.

1 Introduction

EEG contains neural activity alongside ocular and myogenic artifacts, making denoising necessary for neural analysis. Deep-learning denoising is constrained by limited ground-truth clean EEG data and standardized benchmarks, which EEGdenoiseNet addresses with a structured dataset and baseline networks.

  • Motivation: EEG recordings combine brain activity with ocular and myogenic artifacts that contaminate neural signals.Ocular artifacts often appear as large frontal pulses, whereas myogenic artifacts frequently affect temporal and occipital regions across a wide frequency spectrum.
  • Prior approaches: Traditional EEG denoising includes regression, adaptive filtering, and blind source separation approaches.Regression subtracts estimated noise from EEG, while adaptive filters dynamically estimate coefficients from the input signal.
  • Deep-learning context: Deep neural networks require large datasets and gold-standard clean EEG to train and evaluate supervised denoising models.Their performance fundamentally depends on the size of training and test datasets.
  • Research gap: The lack of ground-truth clean EEG data and benchmarks has limited development of deep-learning methods for EEG denoising.
  • Study contribution: EEGdenoiseNet provides 4514 clean EEG segments, 3400 ocular artifact segments, 5598 myogenic artifact segments, and four benchmark networks.The networks are a fully-connected network, simple CNN, complex CNN, and RNN trained in a supervised end-to-end fashion.

2 EEGdenoiseNet Dataset

EEGdenoiseNet assembles standardized EEG, EOG, and EMG segments from public repositories and supports controlled synthesis of contaminated EEG with clean ground truth. The dataset uses two-second segments, expert inspection, and a linear mixing procedure whose artifact level is controlled by λ.

  • Data sources: The dataset is built from publicly available EEG, EOG, and EMG repositories to support EEG denoising research.The source studies had local ethical approval and followed the Helsinki Declaration.
  • Preprocessing: All three signal categories are preprocessed, segmented into two-second windows, rescaled, standardized, and visually inspected by an expert.The two-second length balances recovery of temporal and spectral characteristics with the difficulty of obtaining longer artifact-free EEG segments.
  • EEG segments: The EEG component includes 52 participants performing real and imagined hand-movement tasks with 64-channel recordings.Signals were recorded at 512 Hz, filtered, notched, resampled to 256 Hz, and processed with ICLabel to obtain clean ground truth.
  • Artifact segments: Ocular segments come from horizontal and vertical EOG recordings filtered between 0.3 and 10 Hz and segmented into two-second one-dimensional signals.
  • Artifact segments: Facial EMG supplies myogenic segments, which are filtered between 1 and 120 Hz and resampled to 512 Hz because their signal is concentrated at high frequencies.
  • Dataset contents: The dataset contains 4514 EEG, 3400 ocular-artifact, and 5598 myogenic-artifact segments stored in Matlab and Python formats.
  • Signal synthesis: Contaminated signals are generated as y = x + λ · n, where x is clean EEG, n is an ocular or myogenic artifact, and λ controls contamination strength.Lower λ produces higher SNR because fewer artifacts are added; lower SNR corresponds to higher noise.
  • Signal synthesis: The clean EEG x serves as ground truth and the contaminated signal y serves as the input for end-to-end denoising models.

3 Benchmarking deep learning algorithms

The study provides four standard deep-learning networks and validation procedures as benchmark algorithms for EEG denoising. Their evaluation metrics are intended to support comparisons among newly developed denoising methods.

  • Benchmarking: Four standard deep-learning neural networks are trained and validated as benchmark algorithms for EEG denoising.The benchmark evaluation metrics support performance comparisons across EEG denoising algorithms.

3.1 Generating semi-synthetic data

Semi-synthetic contaminated EEG is generated by mixing clean EEG segments with ocular or myogenic artifact segments. The construction uses predefined train, validation, and test splits and multiple signal-to-noise ratio levels.

  • Ocular artifacts: Ocular-contaminated signals combine 3400 EEG segments with 3400 ocular artifact segments using 80% training, 10% validation, and 10% test splits.Random linear mixing is performed at ten SNR levels from -7dB to 2dB.
  • Myogenic artifacts: Myogenic-contaminated signals combine 4514 EEG segments with 5598 myogenic artifact segments to generate training, validation, and test data.EEG segments are upsampled to 512 Hz and some are randomly reused to match the artifact-segment count.

3.2 Network architectures

The benchmark suite includes fully connected, convolutional, and recurrent architectures for mapping contaminated EEG to denoised signals. The complex CNN adds residual skip-layer connections to support deeper feature extraction, while the RNN uses LSTM cells to model long-term dependencies.

  • Fully-connected network: The FCNN benchmark uses four ReLU hidden layers, with layer width matching the input segment length and dropout regularization.The width is 512 samples for ocular artifact reduction and 1024 samples for myogenic artifact reduction.
  • Simple convolution network: The simple CNN contains four one-dimensional convolutional layers with 1×3 kernels, stride 1, and 64 feature maps.Each convolutional layer is followed by batch normalization and ReLU activation; flattening and a dense layer reconstruct the signal.
  • Complex convolution network: The complex CNN is a one-dimensional residual CNN whose skip-layer connections help avoid gradient explosion during deeper-network training.The residual structure is intended to improve feature extraction relative to the simple CNN.
  • Recurrent neural network: The recurrent benchmark is an LSTM network that sequentially processes each EEG sample and obtains outputs from the state of each cell.Its output network has three fully connected ReLU layers with dropout and 512 or 1024 neurons per layer.

3.3 Learning process

The networks learn an end-to-end nonlinear mapping from normalized contaminated EEG segments to denoised segments, using ground-truth clean EEG and mean squared error optimization. Input normalization uses the contaminated segment's standard deviation, which is retained to restore output magnitude.

  • Normalization: The contaminated and ground-truth EEG segments are normalized by dividing by the contaminated segment's standard deviation.The normalization is intended to facilitate learning.
  • Normalization: The saved standard deviation of each noise segment is multiplied by the network output to restore the denoised EEG segment's magnitude.This reverses the scale introduced during input normalization.
  • End-to-end mapping: End-to-end training maps a normalized contaminated EEG segment directly to a denoised EEG segment through a learned nonlinear function.The function maps contaminated EEG ˆy to denoised EEG ˜x.
  • Optimization: Mean squared error is minimized with gradient descent to reduce the difference between the denoised segment and the ground-truth clean segment.The loss is computed over temporal samples of the segment.
  • Optimization: Training uses artifact-specific epoch schedules and the Adam optimizer with α = 5e−5, β1 = 0.5, and β2 = 0.9.Epoch counts differ across architectures and between ocular and myogenic artifact removal.
  • Implementation: All four networks are implemented in Python 3.7 with TensorFlow 2.2 and trained using two NVIDIA Tesla V100 GPUs.Benchmarking code is publicly available online.

3.4 Performance Evaluation as Benchmarks

Performance evaluation combines convergence analysis, temporal and spectral reconstruction errors, and correlation with clean EEG. The study also compares deep learning methods with empirical mode decomposition and fixed filtering baselines.

  • Convergence: Network convergence is assessed from averaged training and validation loss as a function of training epochs.The curves provide information about the learning procedure and generalization ability.
  • Quantitative metrics: Three objective measures evaluate denoised data: temporal RRMSE, spectral RRMSE, and correlation coefficient.These metrics quantify reconstruction error in time and frequency domains and similarity with the clean signal.
  • Quantitative metrics: Spectral evaluation uses power spectral density over the 0–120 Hz frequency range, with FFT length equal to the input segment length.The PSD function is applied to an input segment to calculate spectral-domain error.
  • Traditional baselines: Traditional comparisons use empirical mode decomposition and filtering, with high-pass filtering at 12 Hz for ocular artifacts and band-pass filtering from 12–40 Hz for myogenic artifacts.The traditional methods are tested 10 times on randomly generated datasets.

4 Results

Across ocular and myogenic artifact removal, deep learning methods generally outperformed traditional methods, especially under heavier noise, while model performance varied by artifact type and architecture. Qualitative results showed strong artifact attenuation and EEG reconstruction, but residual high-frequency noise and myogenic-artifact overfitting remained.

  • Qualitative examples showed substantial artifact attenuation and well-reconstructed noise-free EEG in both ocular and myogenic removal.Frequency-domain results indicated that low-frequency artifacts were attenuated, while high-frequency components retained residual noise.
  • CNNs appeared to overfit during myogenic artifact removal, as their training losses decreased faster while validation losses increased.
  • Deep learning methods achieved lower RRMSE and higher CC than traditional methods for both ocular and myogenic artifact removal, with larger differences at low SNR.At higher SNR, the performance gap between deep learning and traditional methods decreased.
  • RNN performed best for ocular artifact removal, whereas the complex CNN performed best for myogenic artifact removal across the reported benchmark metrics.The comparison used RRMSEtemporal, RRMSEspectral, and CC across multiple SNR levels.
  • DL-based methods significantly outperformed traditional methods for ocular artifact removal on RRMSEtemporal, RRMSEspectral, and CC, with p < 0.001 for each metric.
  • For ocular artifact removal, the mixed artifacts increased delta and theta power ratios and reduced ratios in the other frequency bands.The simple CNN most closely matched ground-truth delta, theta, and beta power ratios; other networks were closest in selected bands.

5 Discussion

EEGdenoiseNet addresses limited standardized resources for deep-learning EEG denoising by combining a structured dataset with benchmark algorithms and evaluation measures. The authors find that deep-learning denoising is feasible, while performance varies with signal quality, artifact frequency, and dataset scope.

  • Dataset and benchmarking: EEGdenoiseNet provides pre-processed EEG, ocular-artifact, and myogenic-artifact segments for generating semi-synthetic noisy data with clean ground truth.The dataset is intended to reduce the electrophysiological and signal-processing preparation required for deep-learning denoising.
  • Dataset and benchmarking: The benchmark combines a standardized EEG dataset with four neural networks and evaluation using RRMSE, PSD, CC, and convergence by loss across epochs.The networks include an FCNN, simple CNN, complex CNN, and RNN.
  • Benchmark findings: Deep-learning networks demonstrated feasibility for artifact attenuation and outperformed the two compared traditional methods, EMD and filter.For ocular artifacts, network CC values were comparable to previously reported regression- and offline-ICA-based methods; myogenic-artifact RRMSEtemporal values were comparable to prior ICA- and canonical-correlation-based methods.
  • Benchmark findings: Neural networks performed better on high-SNR signals than low-SNR signals, while high-frequency artifacts such as EMG were more difficult to remove.The authors relate this pattern to the F-Principle, under which networks learn low-frequency information earlier and high-frequency information later during training.
  • Practical implications: Deep-learning denoising requires substantial ground-truth data but can then be applied flexibly to new multichannel or task-related EEG without corresponding artifact-reference channels.The authors also describe hierarchical representations as enabling direct learning from contaminated inputs to cleaned EEG outputs.
  • Limitations and future scope: The dataset is limited by its data diversity, 2-second segment length, and single-channel focus, and it currently emphasizes ocular and myogenic artifacts.The authors propose adding diverse EEG and artifact types, extending the resource to continuous EEG, and supporting multichannel denoising.

6 Conclusion

EEGdenoiseNet is a publicly available, well-structured benchmark resource containing clean EEG and artifact segments alongside evaluation tools. Benchmarking suggests that deep-learning methods can remove ocular and myogenic artifacts even at high noise levels.

  • Conclusion: EEGdenoiseNet contains thousands of clean EEG, ocular-artifact, and muscular-artifact segments for benchmarking deep-learning EEG denoising methods.The dataset is publicly available in different formats and is accompanied by benchmark tools for evaluating newly developed models.
  • Conclusion: Benchmarking results suggested that deep-learning methods have great potential to remove ocular and myogenic artifacts even at high noise levels.The authors state that the study may accelerate development of deep-learning-based EEG denoising.

Figures and Tables

The figures and tables depict EEGdenoiseNet’s data-generation pipeline, supervised neural-network framework, benchmark architectures, qualitative examples, learning curves, and performance across SNR levels.

  • Dataset pipeline: Figure 1 shows preprocessing of EEG, EOG, and EMG, followed by segment collection and semi-synthetic mixing of EEG with artifacts.The dataset is stored in .mat and .npy formats.
  • Dataset examples: Figure 2 presents EEG, vertical and horizontal EOG, and EMG examples using time-course plots above and PSD plots below.
  • Denoising framework: Figure 3 depicts contaminated one-dimensional segments entering supervised end-to-end networks, which output cleaned EEG evaluated with multiple metrics for ocular and myogenic artifacts.
  • Benchmark networks: Figure 4 compares FCNN, simple CNN, complex CNN, and RNN architectures, with inputs and outputs sized 1 × 512 for ocular and 1 × 1024 for myogenic artifact removal.
  • Qualitative performance: Figures 5 and 6 show best- and worst-case denoising examples in temporal and spectral domains for ocular and myogenic artifact removal.Orange, green, and blue lines denote ground-truth EEG, noisy EEG, and cleaned EEG, respectively.
  • Training behavior: Figure 7 plots training and validation MSE loss against epochs separately for ocular and myogenic artifact removal.Red denotes training and blue denotes validation learning curves.
  • Benchmarks and tables: Figure 8 compares four deep-learning networks with traditional methods across SNR levels for ocular and myogenic artifact removal, while Tables 1–3 summarize collections and frequency-band power ratios.Figure 9 identifies EMD and filtering as the traditional comparison methods.
Loading 2009.11662v4…