Adaptive Thresholding in EEG Artifact Removal Through Multimodal Fusion: A Multimodal Artifact Subspace Reconstruction Approach

Abstract

The removal of artifacts is essential for improving the quality and reliability of electroencephalogram (EEG) data in academic research. Traditional methods, such as mix source separation and signal space projection, often involve subjective and time-consuming manual parameter selection, which is ineffective for artifacts closely correlated with EEG signals. Furthermore, existing artifact removal methods are difficult to generalize across different datasets and experimental conditions. Although artifact subspace reconstruction shows promise, it remains computationally complex and sensitive to parameter selection, limiting its real-time applicability and ability to handle complex artifacts. This study proposes the Multimodal Artifact Subspace Reconstruction (MASR) method, which reduces manual intervention and improves automatic detection and removal of complex artifacts. MASR proposes a new use of multimodal feature extraction techniques, innovatively providing an informative reference for processing EEG signals to reduce artifacts across channels. MASR enhances artifact removal by introducing a novel channel significance metric for quantifying artifact contamination and employing a dynamic adaptive threshold to reduce parameter dependency. MASR integrates multimodal features through principal component analysis (PCA) and ensures cross-modal consistency with Pearson correlation coefficient (PCC) for EEG artifact removal, solving the challenge of artifact characteristics. The MASR method offers a robust, data-driven solution that improves the quality and reliability of EEG data across various applications.