Inter-Trial Analysis of Post-Movement Beta Activities in EEG Signals Using Multivariate Empirical Mode Decomposition

Abstract

Event-related desynchronization/synchronization (ERD/ERS) is a technique to quantify subject's nonphase-locked neural activities underlying specific frequency bands, reactive to external/internal stimulus. However, conventional ERD/ERS studies usually utilize fixed frequency band determined from one or few channels to filter whole-head EEG/MEG data, which may inevitably include task-unrelated signals and result in underestimation of reactive oscillatory activities in multichannel studies. In this study, we adopted multivariate empirical mode decomposition (MEMD) to extract beta-related oscillatory activities in performing self-paced right and left index-finger lifting tasks. The MEMD extracts common modes from all channels in same-index intrinsic mode functions (IMFs) which allows the temporal-frequency features among different channels can be compared in each subband. The beta-band oscillatory activities were further bandpass filtered within trial-specific beta bands determined from sensorimotor-related channels (C3 and C4), and then rectified using amplitude modulation method to detect trial-by-trial beta rebound (BR) values in ERS time courses. The validity of the MEMD approach in BR values extraction has been demonstrated in multichannel EEG study which showed larger BR values than conventional ERS technique. The MEMD-based method enables the trial-by-trial extraction of sensorimotor oscillatory activities which might allow the exploration of subtle brain dynamics in future studies.

Comparison between traditional FFT and marginal spectra using the Hilbert-Huang transform method for the broadband spectral analysis of the EEG in healthy humans

“…Slightly reduced values of the relative PSD expressed in normalized units (%) in the alpha range considered in this study (8-13 hz), observed for the marginal spectrum method, were the result of a better detection using this technique of higher values in other ranges of the spectra, and particularly the increase in the beta1 and gamma fast frequency bands, and in the theta band. In the last years the study of the beta band EEG activity has been the object of motor imagery classification study for brain computer interfaces and assessment of movement of the fingers (Chang et al 2013;Kim et al 2016;Park et al 2013) using the EMD and the MEMD, and it has been shown that the spectra of the IMFs corresponding to the beta frequencies give a better quantitative information than the FFT method applied directly to the EEG original signal. This fact has been also confirmed for the detection of gamma band activity during motor tasks (Amo et al 2017), to assess the gamma activity in the coordination of spatially directed limb and eye movements (Park et al 2014), and to demonstrate the increase of the gamma band power triggered by the anesthetic Ketamine (Tsai et al 2016).…”

Sparsity-based correction of exponential artifacts

“…MEMD is a recently developed algorithm extending conventional EMD. It has been used in different aspects of EEG signal analysis and applications [9,50,49], including removing the ocular artifacts (OA) from multichannel EEG data [53,42]. In this example, we use 4 EEG channels (Fp1, Fp2, C3, C4) measured simultaneously as the input, and decompose the higher-index IMFs (low-frequency subbands) considered to be artifacts [53, Section V].…”

Learning Target Template for Acoustic Event Detection From Low-SNR Training Data

“…The noise disturbance has been considered a crucial issue that affects feature extraction and detection of fault signals in mechanical systems [13]. In other event detection areas like brain oscillatory activities electroencephalography (EEG) [14]- [15], useful signals are buried in EEG recordings and to locate these transients is a great challenge. In these circumstances, the events of interest suffer from lack of information which is twofold: 1) there are only a limited number of target events for training; 2) the target events in the training data are buried in strong noise, which is a low-SNR scenario.…”