Enhancing EEG artifact removal through neural architecture search with large kernels

Abstract

Electroencephalography (EEG) stands as one of the most vital noninvasive tools in neuroscience and clinical practice. Nevertheless, EEG data is highly susceptible to interference from various artifacts, which, in turn, can severely impact the subsequent analyses. As a result, the removal of these unwanted artifacts is of utmost importance. Recently, deep learning methods have demonstrated superior performance in artifact removal compared to traditional approaches. However, experts often invest substantial time and effort in identifying an efficient architecture, a process that is time-consuming and labor-intensive. In light of this challenge, this study introduces, for the first time, an artifact removal method based on neural network architecture search. This approach assigns probabilities to each potential operation within the network and optimizes the most suitable architecture based on the characteristics of the input data. Additionally, we expand the search space by incorporating large convolutional kernels, enabling the network to encompass a wider receptive field for the more effective capture of inherent EEG characteristics. The proposed method is evaluated on publicly available datasets, including electromyography (EMG), electrooculogram (EOG), electrocardiogram (ECG), and motion artifacts. Our results demonstrate that architectures incorporating convolution operations with varied kernel scales and shortcut connections are particularly effective for artifact removal. Notably, our method outperforms state-of-the-art techniques, achieving an average correlation coefficient (CC) exceeding 0.95, a relative root mean squared error (RRMSE) below 0.3, and a signal-to-noise ratio (SNR) above 12 dB. These findings underscore the potential of our method as a reliable and advanced technique for EEG denoising.