Denosieformer: A Transformer-Based Approach for Single-Channel EEG Artifact Removal

As an essential tool in clinical medicine, brain research, and the study of neurological diseases, the electroencephalogram (EEG) is susceptible to various physiological signals, posing difficulties for subsequent research and analysis. Deep learning (DL), an end-to-end method without expert knowledge, is gaining attention in the field of EEG artifact removal. However, the existing DL methods cannot globally observe the latent relationship of EEG signals. They lack the capacity to effectively explore the feature patterns of the signals, resulting in unsatisfactory denoising performance. To address the aforementioned issues, this article proposes a transformer-based EEG denoising architecture (denoiseformer) to solve the artifact removal problem of single-channel contaminated EEG signals. First, the contaminated EEG signals are divided into multiple slices as input so that the transformer module can extract the potential pattern relationships between the slices. Then, we design a multiscale feature extraction and fusion module to obtain high-resolution features with richer and more robust information. Next, a novel attention mechanism, named slice pattern attention, is applied to extract the global information on the temporal dimension of each EEG slice. In addition, a residual variational autoencoder structure can generate superior latent codes between different EEG samples, which help the encoder to distinguish contaminated signal patterns, and hence, improve the recovery performance for artifacts. Comprehensive experiments on three types of semisimulated and one real-laboratory artifacts show our model can obtain excellent performance on two quantitative metrics (relative root-mean-square error and correlation coefficient) for artifact removal tasks in various scenarios.