Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Removing Neural Signal Artifacts with Autoencoder-Targeted Adversarial Transformers (AT-AT)

Benjamin J. Choi

TL;DR

EMG noise significantly contaminates EEG signals, limiting non-invasive BCIs. The authors introduce AT-AT, a compact autoencoder-guided transformer denoising framework that uses a correlation-based noise proxy to selectively apply a time-series transformer and employs adversarial training to preserve EEG spectral properties. Evaluated on semi-synthetic EEGdenoiseNet data with 67 subjects, AT-AT achieves high reconstructive accuracy (CC > 0.95 at 2 dB, CC ≈ 0.70 at -7 dB) while reducing model size by over 90% and showing competitive performance versus larger models, with corroborating real-world downstream benefits. The approach demonstrates practical potential for efficient EEG artifact removal in resource-constrained BCIs, including downstream improvements in digit-versus-non-digit classification on MindBigData.

Abstract

Electromyogenic (EMG) noise is a major contamination source in EEG data that can impede accurate analysis of brain-specific neural activity. Recent literature on EMG artifact removal has moved beyond traditional linear algorithms in favor of machine learning-based systems. However, existing deep learning-based filtration methods often have large compute footprints and prohibitively long training times. In this study, we present a new machine learning-based system for filtering EMG interference from EEG data using an autoencoder-targeted adversarial transformer (AT-AT). By leveraging the lightweight expressivity of an autoencoder to determine optimal time-series transformer application sites, our AT-AT architecture achieves a >90% model size reduction compared to published artifact removal models. The addition of adversarial training ensures that filtered signals adhere to the fundamental characteristics of EEG data. We trained AT-AT using published neural data from 67 subjects and found that the system was able to achieve comparable test performance to larger models; AT-AT posted a mean reconstructive correlation coefficient above 0.95 at an initial signal-to-noise ratio (SNR) of 2 dB and 0.70 at -7 dB SNR. Further research generalizing these results to broader sample sizes beyond these isolated test cases will be crucial; while outside the scope of this study, we also include results from a real-world deployment of AT-AT in the Appendix.

Removing Neural Signal Artifacts with Autoencoder-Targeted Adversarial Transformers (AT-AT)

TL;DR

EMG noise significantly contaminates EEG signals, limiting non-invasive BCIs. The authors introduce AT-AT, a compact autoencoder-guided transformer denoising framework that uses a correlation-based noise proxy to selectively apply a time-series transformer and employs adversarial training to preserve EEG spectral properties. Evaluated on semi-synthetic EEGdenoiseNet data with 67 subjects, AT-AT achieves high reconstructive accuracy (CC > 0.95 at 2 dB, CC ≈ 0.70 at -7 dB) while reducing model size by over 90% and showing competitive performance versus larger models, with corroborating real-world downstream benefits. The approach demonstrates practical potential for efficient EEG artifact removal in resource-constrained BCIs, including downstream improvements in digit-versus-non-digit classification on MindBigData.

Abstract

Electromyogenic (EMG) noise is a major contamination source in EEG data that can impede accurate analysis of brain-specific neural activity. Recent literature on EMG artifact removal has moved beyond traditional linear algorithms in favor of machine learning-based systems. However, existing deep learning-based filtration methods often have large compute footprints and prohibitively long training times. In this study, we present a new machine learning-based system for filtering EMG interference from EEG data using an autoencoder-targeted adversarial transformer (AT-AT). By leveraging the lightweight expressivity of an autoencoder to determine optimal time-series transformer application sites, our AT-AT architecture achieves a >90% model size reduction compared to published artifact removal models. The addition of adversarial training ensures that filtered signals adhere to the fundamental characteristics of EEG data. We trained AT-AT using published neural data from 67 subjects and found that the system was able to achieve comparable test performance to larger models; AT-AT posted a mean reconstructive correlation coefficient above 0.95 at an initial signal-to-noise ratio (SNR) of 2 dB and 0.70 at -7 dB SNR. Further research generalizing these results to broader sample sizes beyond these isolated test cases will be crucial; while outside the scope of this study, we also include results from a real-world deployment of AT-AT in the Appendix.

Paper Structure

This paper contains 6 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Results & Discussion
  4. Appendix
  5. Additional Model Specifics
  6. Post-Processing for BCI Deployment on the Imagined Digit Classification Problem

Figures (4)

  • Figure 1: The autoencoder-targeted adversarial transformer (AT-AT) system architecture. The initial signal is passed to the LSTM-CNN (LC) targeting model and denoising autoencoder (upper panels); the transformer is subsequently invoked on select target sites (lower diagram). Design incorporates elements featured in Choi2025bChoi2025c.
  • Figure 2: AT-AT model workflow. From top to bottom: (1) raw input signal, (2) initial autoencoder filtration pass with high-noise target site masking, (3) time-series transformer-based reconstruction of target sites, (4) ground truth signal.
  • Figure 3: AT-AT performance relative to major deep learning benchmarks. [A], [B], and [C] refer to Zhang2021ICASSPYin2023 and Cui2024, respectively. All values are best extrapolated from reported documentation (best-in-class values in bold).
  • Figure 4: AT-AT processing enables separability between neural signal classes (i.e., digit- vs. non-digit-related thoughts) to be elucidated. Pre-AT-AT t-SNE is depicted left, while the post-AT-AT t-SNE is depicted right; also given in Choi2025b.