All research

Physiological signal processing · 2022–2024

Deep learning restoration of motion-corrupted EEG

Deep learning methods that remove motion artefacts from single-channel EEG while preserving the signal in time and frequency.

Qatar National Research Fund NPRP12S-0227-190164 · Qatar University IRCC-2021-001

Block diagram. Motion-corrupted and reference EEG are preprocessed and divided into training, validation and test sets. Deep learning models are trained and tuned, and the best model predicts clean EEG for performance evaluation.
Figure 1 from Mahmud et al. (2023). Framework for correcting motion artefacts in EEG with deep learning.View full-size figureSource articleCC BY 4.0

Approach

Can clean EEG be recovered from a single channel corrupted by electrode motion?

Head and electrode movement can add artefacts far stronger than the EEG itself, which limits recording outside controlled settings. This project first developed MLMRS-Net, a multi-resolution network for reconstructing clean EEG. It then developed AGO-CycleGAN, an attention-guided operational CycleGAN that restores corrupted EEG and recovers its frequency content.

  1. Record motion-corrupted EEG
  2. Restore the EEG waveform
  3. Evaluate time and frequency recovery

Removing motion artefacts with MLMRS-Net

Neural Computing and Applications · 2023

The study used a public PhysioNet dataset of 23 single-channel prefrontal EEG recordings. In each 9-minute recording, two electrodes sat 30 mm apart. One stayed undisturbed as the reference, and the other was tapped for 10 to 25 seconds about every two minutes. Signals were resampled to 256 Hz, baseline-corrected, notch-filtered at 50 Hz and split into 1,024-sample segments.

MLMRS-Net is a five-level, one-dimensional U-shaped network with multi-resolution pooling blocks in its encoder and decoder, and deep supervision at each decoder level. It was compared with 10 other 1D networks, testing each recording in turn on a model trained with the other 22. MLMRS-Net gave the highest artefact reduction and the lowest reconstruction error, while signal-to-noise gains were close across all networks.

EEG alone reached 89.32% artefact reduction. Adding three-axis accelerometer signals raised this to 90.52%, so strong restoration remained possible without extra motion sensors.

Motion artefact reduction
90.52%
Signal-to-noise improvement
26.64 dB
Mean absolute error
0.056
Selected networks from Table 1 of Mahmud et al. (2023), each using EEG and three-axis accelerometer inputs. Results are averaged across 23 test folds.
Signal-to-noise improvementArtefact reductionMean absolute error
FPN26.489 dB86.197%0.069
U-Net26.568 dB88.817%0.064
U-Net++26.703 dB89.956%0.059
Attention U-Net++26.632 dB90.191%0.058
MLMRS-Net26.641 dB90.516%0.056
U-shaped network diagram. A four-channel input of corrupted EEG and three accelerometer axes passes through five encoder levels and matching decoder levels, with a deep supervision output at each decoder level.
Figure 2 from Mahmud et al. (2023). The MLMRS-Net architecture.View full-size figureSource articleCC BY 4.0
Six groups of EEG plots, each showing the reference, input and MLMRS-Net output. In four groups the corrupted input is heavily distorted and the output follows the reference. In two clean groups all three signals match closely.
Figure 7 from Mahmud et al. (2023). Reference, input and restored EEG for motion-corrupted segments (a–d) and clean segments (e, f).View full-size figureSource articleCC BY 4.0

Restoring EEG with an attention-guided operational CycleGAN

Engineering Applications of Artificial Intelligence · 2024

The second study used the same recordings and restored EEG directly from the corrupted channel. Baseline correction was applied to the reference channel, so the model learned to remove baseline drift together with motion artefacts.

AGO-CycleGAN uses attention-guided Feature Pyramid Network generators and PatchGAN discriminators built with Self-ONN layers. Among six CycleGAN variants, this design gave the best temporal correlation, spectral correlation and signal-to-noise improvement. Only the restoration generator is needed after training, with 2.245 million parameters.

Motion-corrupted segments moved close to the reference after restoration, while clean segments stayed almost unchanged.

Signal-to-noise improvement
26.50 dB
Temporal correlation improvement
87.2%
Mean absolute error
0.071
CycleGAN variants from Table 2 of Mahmud et al. (2024). Correlations compare restored EEG with the reference, and parameters cover the full CycleGAN.
ParametersTemporal correlationSpectral correlationSignal-to-noise improvement
CNN U-Net2.058 M0.8480.91825.619 dB
CNN FPN2.062 M0.8570.92125.210 dB
CNN attention FPN2.076 M0.8690.93326.328 dB
Self-ONN U-Net5.402 M0.8650.92625.873 dB
Self-ONN FPN5.408 M0.8760.93826.302 dB
AGO-CycleGAN5.540 M0.8800.94326.497 dB
CycleGAN framework diagram. One generator restores motion-corrupted EEG and a second generator maps clean EEG to corrupted EEG. Two discriminators, cycle-consistency losses and identity losses link the two directions.
Figure 3 from Mahmud et al. (2024). The AGO-CycleGAN training framework.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
Architecture diagrams of the Self-ONN generator with attention gates on its skip connections, the PatchGAN discriminator, and the guided attention block.
Figure 4 from Mahmud et al. (2024). The generator, discriminator and guided attention block.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
Six groups of EEG plots, each showing the reference, input and AGO-CycleGAN output. Corrupted inputs with motion and baseline drift are restored towards the reference, and clean inputs remain nearly unchanged.
Figure 5 from Mahmud et al. (2024). Reference, input and restored EEG for corrupted segments (a–d) and clean segments (e, f).View full-size figureSource article© 2023 Elsevier Ltd · author reuse

Recovering the EEG frequency profile

Engineering Applications of Artificial Intelligence · 2024

Motion artefacts added strong low-frequency power. The delta band rose from 10.7% to 60.1% of total EEG power, while the alpha band fell from 53.2% to 19.1%. After restoration, delta and alpha returned to 11.9% and 50.4%, close to the reference.

Spectral correlation with the reference increased from 0.128 to 0.943. Power spectra and topographic maps showed the same recovery across the EEG frequency bands.

Spectral correlation improvement
93.5%
Spectral correlation after restoration
0.943
Restored delta power (reference 10.7%)
11.9%
Share of EEG power in each frequency band across the dataset, from Table 1 of Mahmud et al. (2024).
DeltaThetaAlphaBetaGamma
Reference0.10700.20700.53170.11150.0430
Motion-corrupted0.60110.12040.19130.05840.0290
Restored0.11880.20590.50400.11720.0542
Power spectral density plot from 0 to 40 Hz. The motion-corrupted EEG lies far above the reference across all frequencies, while the restored EEG almost overlaps the reference.
Figure 6 from Mahmud et al. (2024). Power spectral density of reference, motion-corrupted and restored EEG.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
Grid of scalp topographic maps for the whole spectrum and five EEG bands. Maps of motion-corrupted EEG show strong prefrontal power, especially in the delta band, while restored maps resemble the reference maps.
Figure 7 from Mahmud et al. (2024). Topographic maps of reference, motion-corrupted and restored EEG across frequency bands.View full-size figureSource article© 2023 Elsevier Ltd · author reuse

My contribution

I led the MLMRS-Net and AGO-CycleGAN studies, including model development, implementation, experiments, quantitative and qualitative analysis, and manuscript preparation. I also contributed to the studies in this project on eye and muscle artefact removal.

Study scope

Both studies used the same public dataset of 23 single-channel prefrontal recordings, with motion created by tapping one electrode and a nearby electrode providing the reference. Validation during natural movement, on multi-channel EEG and without a clean reference is still needed.