A multi-class EEG-based BCI classification using multivariate empirical mode decomposition based filtering and Riemannian geometry

• Published in: Expert systems with applications Vol. 95; pp. 201 - 211
• Main Authors: Gaur, Pramod, Pachori, Ram Bilas, Wang, Hui, Prasad, Girijesh
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.04.2018; Elsevier BV
• Subjects: Covariance matrix; Electroencephalography; Filtration; Human motion; Human-computer interface; Image classification; Image enhancement; Interfaces; Multichannel communication; Signaling
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2017.11.007

•A subject specific MEMD based filtering is proposed to classify MI EEG signals.•The mean frequency is used to filter the MIMFs to obtain enhanced EEG signals.•The sample covariance matrix feature is extracted from these enhanced EEG signals.•This feature is classified using Riemannian geometry.•Our combined approach outperformed other recognition methods. A brain-computer interface (BCI) facilitates a medium to translate the human motion intentions using electrical brain activity signals such as electroencephalogram (EEG) into control signals. EEG signals are non-stationary and subject specific. A major challenge in BCI research is to classify human motion intentions from non-stationary EEG signals. We propose a novel subject specific multivariate empirical mode decomposition (MEMD) based filtering method, namely, SS-MEMDBF to classify the motor imagery (MI) based EEG signals into multiple classes. The MEMD method simultaneously decomposes the multichannel EEG signals into a group of multivariate intrinsic mode functions (MIMFs). This decomposition enables us to extract the cross-channel information and also localize the specific frequency information. The MIMFs are considered as narrow-band, amplitude and frequency modulated (AFM) signals. The statistical measure, mean frequency has been used to automatically filter the MIMFs to obtain enhanced EEG signals which better represent motor imagery related brainwave modulations over μ and β rhythms. The sample covariance matrix has been computed and used as a feature set. The feature set has been classified into multiple MI tasks using Riemannian geometry. The proposed method has helped achieve mean Kappa value of 0.60 across nine subjects of the BCI competition IV dataset 2A which is superior to all the reported methods.