Transformer-Based Autoencoder for Denoising EEG and ECoG Signals: A Breakthrough in Neural Signal Processing

Electroencephalography (EEG) and electrocorticography (ECoG) signals are pivotal for brain-computer interface (BCI) technologies, yet their utility is hindered by pervasive noise from physiological artifacts and environmental sources. This study introduces a novel Transformer-based autoencoder that revolutionizes denoising of EEG and ECoG signals, achieving unprecedented signal-to-noise ratios (SNRs) of 15.32 dB for EEG and 12.45 dB for ECoG, markedly surpassing the Independent Component Analysis (ICA) baseline of 8.76 dB and 7.89 dB, respectively. Leveraging the PhysioNet EEG Motor Movement/Imagery Dataset with synthetic ECoG noise, our approach harnesses the Transformer's attention mechanisms to model complex temporal dependencies, offering a fully automated, scalable solution. Comprehensive visualizations and spectral analyses underscore the method's efficacy, positioning it as a new state-of-the-art in neural signal processing with transformative potential for BCI applications, neuroprosthetics, and clinical diagnostics.