Multivariate Variational Mode Decomposition based approach for Blink Removal from EEG Signal

Electroencephalography (EEG) signals contain ocular artifacts which degrades the overall performance of any neuro-engineering based analysis or applications like brain computer interfaces. In general, independent component analysis (ICA) is used for removing blinks. However, that requires expert intervention. This paper aims at cleaning the eye blink related artifacts automatically without any manual interventions. We propose a novel approach based on multivariate extension of variational mode decomposition (VMD), called MVMD, for the said purpose. The mode-alignment property of MVMD has been utilized to align the joint/common oscillations across multiple channels of a given single mode. The detection of blinks is found to be better in the components of MVMD over the raw EEG signal. The proposed approach is first validated on synthetically generated EEG data and then it is tested on two publicly available real EEG datasets. Results confirm usability of the proposed approach over ICA technique. An average correlation of 0.938 (±0.0221) and 0.9869 (±0.0094) are obtained for the synthetically generated and high end EEG data, respectively, in the non-blink regions. We obtained approximately 90% classification accuracy in detecting fatigue on CogBeacon dataset. This accuracy is comparable with that obtained using state of the art approach, with the added advantage of not requiring manual interventions of experts.