TL;DR
DeepSeparator splits an EEG recording into “brain” and “artifact” inside a neural network’s embedding space. It cleans eye-blink and muscle noise better than classic methods, and because it outputs the removed artifact too, you can check its work.
The DeepSeparator network (top), error across noise levels compared with classic methods (bottom left), and an example of eye-blink removal (bottom right). Figures from the paper.
The DeepSeparator network (top), error across noise levels compared with classic methods (bottom left), and an example of eye-blink removal (bottom right). Figures from the paper.

The question

EEG measures tiny brain voltages from the scalp, and it is easily swamped by other electrical activity in the body. Blinks and eye movements (EOG) and muscle tension (EMG) can be many times larger than the brain signal itself.

Classic cleaning tools, such as filters, Hilbert–Huang transforms, and ICA, each depend on assumptions about what the noise looks like. They often need an expert to decide which components to throw away. Deep-learning denoisers are more automatic, but they are black boxes. Can we get automatic cleaning whose decisions we can still inspect?

What we did

DeepSeparator has three parts:

  1. An encoder, built from stacked Inception blocks, turns raw EEG into a rich embedding.
  2. A decomposer learns which parts of that embedding belong to artifacts and turns them down.
  3. A decoder rebuilds clean EEG from what’s left, and it can also rebuild the artifact that was removed.

We trained it on EEGdenoiseNet, a benchmark that mixes clean EEG with recorded eye and muscle artifacts at known noise levels, so the true clean signal is known. We then applied the trained model to real multi-channel EEG from an auditory experiment.

(a) Encoder, decomposer, and decoder. (b) The Inception block. (c) Training uses three kinds of input (noisy EEG, clean EEG, and pure artifact) so the model learns to separate them.
(a) Encoder, decomposer, and decoder. (b) The Inception block. (c) Training uses three kinds of input (noisy EEG, clean EEG, and pure artifact) so the model learns to separate them.

What we found

1. Lower error than classic methods at every noise level

For eye artifacts, DeepSeparator had the lowest error in both time and frequency and the highest correlation with the true clean signal. This held across signal-to-noise ratios from –7 to 2 dB, against adaptive filtering, the Hilbert–Huang transform, and EEMD-ICA. Muscle artifacts showed the same pattern.

EOG artifact removal across noise levels: temporal error (a), spectral error (b), and correlation with ground truth (c). DeepSeparator is shown in red.
EOG artifact removal across noise levels: temporal error (a), spectral error (b), and correlation with ground truth (c). DeepSeparator is shown in red.

2. It shows you what it removed

Because the decoder can output the artifact as well as the clean EEG, you can see exactly what was taken out. In practice, the extracted artifacts look like blinks and muscle bursts. This makes the model much easier to trust than a typical black-box denoiser.

Raw EEG (a), denoised EEG (b), and the artifact DeepSeparator extracted (c), for two eye-artifact and two muscle-artifact examples.
Raw EEG (a), denoised EEG (b), and the artifact DeepSeparator extracted (c), for two eye-artifact and two muscle-artifact examples.

3. It works on real, multi-channel recordings

Although it was trained on single-channel, semi-synthetic data, DeepSeparator transferred to real multi-channel EEG. Like ICA, it removed eye artifacts, but without anyone picking components by hand. It also suppressed high-frequency, muscle-like noise. The paper goes on to compare ERPs and source localization after both kinds of cleaning.

Real EEG: raw signals (a), ICA components and artifacts (b–c), ICA-cleaned EEG (d), and DeepSeparator's extracted artifacts (e) and cleaned EEG (f).
Real EEG: raw signals (a), ICA components and artifacts (b–c), ICA-cleaned EEG (d), and DeepSeparator’s extracted artifacts (e) and cleaned EEG (f).

My take

BibTeX
@article{yu2022deepseparator,
  title   = {Embedding Decomposition for Artifacts Removal in {EEG} Signals},
  author  = {Yu, Junjie and Li, Chenyi and Lou, Kexin and Wei, Chen and Liu, Quanying},
  journal = {Journal of Neural Engineering},
  volume  = {19},
  number  = {2},
  year    = {2022},
  doi     = {10.1088/1741-2552/ac63eb}
}