Relative spectral power (RSP) and temporal RSP as features for movement imagery EEG classification with linear discriminant analysis

Real time control of brain computer interface (BCI) is very attractive, challenging and interesting research areas. For the same many signal processing methods have been applied and different feature extraction methods have been investigated to maximize mutual information (MI), Cohen kappa value, accuracy (ACC) for electro-encephalogram (EEG) based BCI system. This paper presents a novel features extraction method for real time classification of right and left motor imagery for developing EEG based BCI system. A novel features relative spectral power (RSP) and temporal variations of relative spectral power (TRSP) in motor rhythm is proposed and applied to Graze BCI dataset of BCI competitions 2003. We also consider autoregressive (AR) coefficient as features. We apply Burg and Yule-Walker methods of AR coefficient estimation technique. Extracted features is subjected to linear discriminant classifier (LDC), whose design based on linear least square method, Fisher criterion and Mahanlanobis distance based approach, to classify the every time samples of right and left hand movement imagination. The obtained result have been compared with the BCI competitions 2003 results on the same data set in terms of either the mutual information or misclassification rate. AR coefficient estimation with Yule-Walker method outperform the winner of BCI competitions 2003 and obtain the maximum value of MI 0.69 bits at 5.21s while the TRSP obtain maximum value of MI 0.54bits at 5.24s. TRSP outperform the second winner who has maximum MI 0.46 bits at 5.05s.