Nonconvex Sparse Regularization Method for Eyeblink Artifact Suppression From Single-Channel EEG Signals

Abstract

Recent advancements in affordable single-channel electroencephalogram (EEG) devices have garnered considerable attention due to their ability to reduce hardware complexity. However, effectively suppressing eyeblink artifacts in single-channel EEG signals remains a substantial challenge for biomedical applications. This article proposes a nonconvex sparse regularization methodology (NSRM), which explores the generalized minimax-concave (GMC) penalty for eyeblink artifact suppression from single-channel EEG signals. The contaminated EEG signals can be initially modeled within the sparse representation framework as a combination of target and noise components. The proposed methodology preserves the convexity of the sparsity-regularized least square objective function, allowing the global minimum to be reached through convex optimization. Specifically, a forwardbackward splitting (FBS) algorithm is developed to resolve the nonconvex sparse regularization problem of eyeblink artifact suppression. In addition, we introduce an adaptive selection strategy for the regularization parameter. The advantage over conventional methods is that NSRM can better preserve useful information from EEG signals while suppressing eyeblink artifacts. To validate the efficacy of NSRM, a semisimulated EEG dataset and two real experiment datasets have been analyzed. Results demonstrate that our NSRM methodology eliminates eyeblink artifacts effectively and accurately from single-channel EEG signals, outperforming the $L1$ norm-based sparse regularization method, as evidenced by quantitative metrics. Finally, comparison results with the advanced K-means singular value decomposition (K-SVD) have also confirmed the superiority of our proposed NSRM for eyeblink artifact suppression in the context of the sparse representation paradigm.