A Time–Frequency Attention-Based EEG Artifact Removal Network by Interacting and Fusing Artifact Representation

Electroencephalography (EEG) is a vital non-invasive tool in neuroscience and clinical practice, but its signals are often contaminated by various artifacts, such as those from electromyography (EMG), electrooculography (EOG), and electrocardiography (ECG), which degrade the quality of analysis. Although deep learning approaches have demonstrated superior artifact removal compared to traditional methods, they predominantly focus on signal reconstruction, neglecting the frequency-dependent nature of EEG artifacts and the need for effective interaction between artifact and EEG signal components. This limitation restricts their ability to fully eliminate these interferences and preserve relevant EEG features. To address the limitations of existing methods, we propose a novel approach that integrates a multi-conv block for capturing multi-modal information of EEG signals and an improved attention mechanism to effectively focus on both local and global features. Additionally, a frequency enhancement module is introduced to handle the heterogeneous artifact distribution in the frequency domain, while an interactive module facilitates the fusion of artifact and EEG signal components, improving the overall artifact removal process. The method has been validated on various public datasets including EMG, EOG and ECG. In addition, in order to study the practicality of the model, we conducted classification experiments. The research results indicate that the proposed method is highly effective in eliminating artifacts, outperforming existing state-of-the-art methods and achieving an average correlation coefficient of over 0.95, a relative root mean square error of less than 0.26, and a signal-to-noise ratio of over 12 decibels.