No abstract

This work decodes two-class motor imagery (MI) based on four main processing steps: (i) Raw electroencephalographic (EEG) signal is decomposed to single trials and spatial filters are estimated for each trial by common spatial filtering (CSP) method; (ii) features are extracted by taking the log transformation (normal distribution) of the spatially filtered EEG signal; (iii) optimal channel selection algorithm is proposed to reduce the number of EEG channels, such approach is regarded as key technological advantage in the implementation of brain–computer interface (BCI) to reduce the system processing time; (iv) finally, support vector machine (SVM) is employed to discriminate two classes of left and right hand MI. Two variations of SVM were proposed: polynomial function kernel and radial-based function RBF kernel. The results revealed that CSP succeeded in removing the strong correlation bound between the EEG samples by maximizing the variance of class 2 samples while minimizing the variance of class 1 samples. The channel selection algorithm achieved its goal to reduce the data dimension by selecting two channels out of three having the lowest variance entropies of 0.239 and 0.261 for channel 1 and channel 2, respectively. The features vector was divided into 80% train and 20% test with five-fold cross validation. The classification performance of SVM-polynomial kernel was 87.86% while it is 95.72% for SVM-RBF kernel as average accuracy of five-folds for both. Thus SVM-RBF is superior to SVM-Poly in the proposed framework.

In this paper, (i) time domain, frequency domain and spatial domain feature extraction methods were investigated. (ii) Two dimensionality reduction methods were proposed, implemented and compared. (iii) The method pair (feature extraction + dimensionality reduction) that owns the lowest classification error rate will be used to learn a machine learning algorithm to control robotic hand in offline mode. Two classes EEG dataset of three bipolar channels was used. The extracted feature vectors were fed into Support Vector Machine with Radial Basis Function kernel (SVM-RBF) to train the classifier. The experimented time domain feature extraction methods were: Mean Absolute Value (MAV), integrated Absolute Value (IAV), Zero Crossing (ZC), Root Mean Square (RMS), Waveform Length (WL) and Slope Sign Change (SSC). Frequency domain feature was the Autoregressive Feature (AR). Finally, the spatial domain feature was the Common Spatial Patterns (CSP). Matlab codes for Principal Component Analysis (PCA) and channel selection algorithm were designed and used to reduce the dimensionality of the features vector. Results showed that CSP features got the lowest error rate for both dimensionality reduction technique with 2.14%. Results recommends to use channel selection algorithm over PCA since it owns the lowest processing time of 8.2s over 8.5s for PCA.