BCI data classification 2003 - Graz data set Mohan Sadashivaiah ,Akash Narayana, Raveendran Rengaswamy, Shanmukh Katragadda FirstName.secondName@daimlerchrysler.com DaimlerChrysler Research & Technology India Pvt Ltd., 137 Infantry road, Bangalore-560001, India The following feature extraction method was employed for classification of the BCI data set. Training was performed using (training) data between time points from 4s to 6s. A negative classifier output corresponds to class 1, a positive output to class 2, and zero for reject class. AR coefficients for channels 1 and 3 In this method, channels #1 and #3 were used. AR coefficients of order 6 were extracted from both the channels. The AR coefficients were computed for a window of 128 samples with a shift of 1 sample. This results in 12 dimensional feature vector ( 6 AR coefficients from each of the two channels). These features were fed to a LDA, which uses Mahalanobis distance measure for classification. This also results in the classification of the signal at every sample after an initial delay of 128 samples (=1s). Results: The mutual information plot on the training data set is as shown in Fig1. Fig1: Mutual information on the training set [see DOC file].