Computational protocol: EEG resolutions in detecting and decoding finger movements from spectral analysis

Similar protocols

Protocol publication

[…] Datasets recorded at the sampling rate of 1000 Hz were firstly downsampled to 250 Hz to be consistent with other datasets. The first 2-s EEG data in each 6-s trial were removed from further analysis, since the period was designed for subjects to engage unavoidable movements, such as blink or swallowing. The remaining data were then high-pass filtered at 0.3 Hz using an elliptic infinite impulse response (IIR) filter from the EEGLAB toolbox (Delorme and Makeig, ) with both forward and reverse filtering to minimize phase distortions. A 60 Hz notch filter with a transition band of 0.3 Hz was further applied to remove power-line noise. To remove common physiological artifacts, independent component analysis (ICA) (Hyvärinen et al., ) from the EEGLAB toolbox was performed, implemented with the Infomax algorithm (Bell and Sejnowski, ). EEG artifacts, such as generic discontinuities, electrooculogram (EOG), electrocardiogram (ECG), and electromyogram (EMG), were then identified and rejected using the ADJUST toolbox (Mognon et al., ) and visual inspections. Total 64 independent components (ICs) were reconstructed and about 10–20 artifact-related ICs were rejected in each subject.To further increase signal-to-noise ratio (SNR), EEG signals went through a common average reference (CAR) filter (McFarland et al., ), with data from each channel re-referenced to the average of data from all channels. After these steps, the 1-s segments of EEG data in the middle of 2-s period for fixation and 2-s period for movement in each trial were selected for the following analysis. This resulted in total six conditions: EEG data of five finger-movement conditions and pooled EEG data from resting conditions (i.e., fixation periods). It led to 60 ~ 80 one-second movement segments and about five times of resting segments in each subject. [...] Classification procedures for evaluation of spectral features through the two decoding tasks discussed above are described here. Since most EEG features exhibited localized spatial patterns (e.g., mu/beta rhythms over the motor cortex), spectral features from subsets of all channels were used as input features to classifiers to avoid negative impacts from irrelevant channels. For the classic mu/beta rhythms features, channel C3 and its neighboring channels were chosen as feature channels (for right hand movements). For features from the PCs, channels were selected based on r2 values between two compared conditions. In general, channels were ranked by their corresponding r2 values, and then the first 10 channels were chosen as feature channels. If more than two conditions to be compared, i.e., five fingers, the union of selected channels for all finger pairs was used. For cases using combined features, the union of selected channels for each feature was used.The linear support vector machine (SVM) (Vapnik, , ) with radial basis function (RBF), implemented in a MATLAB package, i.e., LIBSVM (Chang and Lin, ), was chosen for classification. The penalty parameter and gamma value in the RBF kernel were determined by a grid-search approach within the range of logarithm value [−10, 20] and [−15, 10], respectively, with the step width of one (Hsu et al., ). The decoding features (projection weights on PCs, and mu/beta PSDs) were linearly scaled into the range [−1, +1] to avoid numeric range dominance of one feature over others. A binary SVM classifier was applied in detecting movements from resting and the one-vs.-one scheme followed by a majority voting was used to solve the multiclass classification problem (Hsu and Lin, ) in decoding five fingers. A five-fold cross validation procedure was implemented, i.e., 80% data for training and 20% data for testing, which was repeated 30 times by randomly partitioning data into training and testing sets. […]

Pipeline specifications

Software tools EEGLAB, LIBSVM
Applications Miscellaneous, Clinical electrophysiology