Classification of non-motor cognitive task in EEG based brain-computer interface using phase space features in multivariate empirical mode decomposition domain. (January 2018)

• Record Type: Journal Article
• Authors: Dutta, Suman; Singh, Mandeep; Kumar, Amod
• Is Part Of: Biomedical signal processing and control. Volume 39(2018)
• Page Start: 378
• Page End: 389
• Publication Date: 2018-01
• Subjects: Electroencephalogram -- Multivariate empirical mode decomposition -- Intrinsic mode function -- Phase space reconstruction -- Singular value decomposition -- Least square support vector machine; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2017.08.004
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Non-motor mental task induces significant changes in the patterns of brain electrical activity. Largest singular value of the phase space trajectory matrix can capture such changes in the MEMD domain. Combination of MEMD and phase space reconstruction provides a fertile ground for nonlinear analysis of EEG signal. The proposed feature extraction approach is in time domain and hence suitable for real time application. Abstract: The purpose of this research paper is to present a new framework for EEG feature extraction based on the combination of multivariate empirical mode decomposition (MEMD) and phase space reconstruction (PSR) for classifying a small set of non-motor cognitive task EEG signals in mental task based multitask brain computer interface(BCI) system. Our proposed approach employed phase space analysis of the intrinsic mode functions (IMFs) generated by MEMD based decomposition of the six channels EEG signals. The combination of two powerful signal processing techniques i.e. MEMD with phase space reconstruction (PSR) enabled us to characterize the nonlinear and non-stationary nature of the dynamics underlying a particular cognitive task more accurately. Our proposed approach consists of three stages, in the first stage; the application of MEMD to multichannel EEG data gave rise to adaptive i.e. data driven decomposition of the multivariate time series data into a set of IMF groups. All the member IMFs within a group have common oscillatory frequency but different amplitude and cortical origin. In the second stage, a small subset of IMF groups was selected according to their task correlation factor and subsequently represented in the two dimensional phase space through their trajectory matrices. In the third stage, largest singular values of the trajectory matrices corresponding to a subset of sensitive IMFs were employed for forming the feature vectors. Finally, the extracted feature vectors were fed to a least square support vector machine (LS-SVM) classifier for binary i.e. pair wise classification of these mental task EEG signals. With the new feature vectors, it is shown that LS-SVM with RBF kernel provides accuracy of 83.33% in classifying between mental arithmetic and mental letter composing. The performance of this classifier was evaluated on various parameters such as accuracy, specificity and sensitivity. The classification results show the potential of the proposed approach for classifying any non-linear a non- stationary signal.