EEG features extraction using PCA plus LDA approach based on L1-norm for motor imaginary classification

Brain-Computer Interfaces (BCIs) are communication systems, in which users use their brain activity instead of original motor movements, to produce signals related to specific intention, which in turn are used to control computers or communication devices attached to it. These activities are generally measured by Electroencephalography (EEG). BCI uses pattern recognition approach in which features are extracted from EEG signals which are used to identify the user's mental state. BCI commonly used feature extraction method is Common Spatial Pattern (CSP). Despite of its effective usefulness, it suffers from intrinsic variations and nonstationarity of EEG data as CSP ignores the within class dissimilarities. Also, the formulation of CSP criteria is based on variance using L2-norm, which makes it sensitive to outliers too. A new PCA plus LDA method based on L1-norm has been proposed alternative to CSP which efficiently considers between the classes and within the class dissimilarities. Also the objective function is reformulated using L1-norm to suppress the effect of outliers. The optimal spatial pattern of given method are obtained by introducing an iterative algorithm. The proposed method was evaluated against Dataset IIa of BCI Competition IV. The result showed that the proposed method outperformed in almost all the cases with low mis-classification rate and results in average kappa value 0.3482.