BCI competiton 2005: Data set IIIb Institute of Biomedical Engineering of Xi'an Jiaotong University Xi'an, 710049, P. R. China Team Member: Xiaomei Pei, Guangyu Bin Supervisor: Prof. Chongxun Zheng 1. Description of our method The NaN samples in train and test dataset are deleted. The spectrum within different subject-specific frequency bands for each 1 second window of 125 samples with a window shift of one sample was computed by FFT, before which the Hanning window was added to reduce spectrum leakage. The spectrum features for of C3 and C4 within different frequency band of each 1 second were constructed as time-varying feature vectors Ft. Classifying by a Fisher Discriminant Analysis, in which the discriminant distance information across time are incorporated for classification. The reason is the brain state should be time continuous in one trial and the discriminant distance reflects the classification confidence. 2. Format of the data The result is saved in a Matlab-fileformat. O3_res.mat, S4_res.mat, X11_res.mat are the classifier output for subject O3, S4, X11 respectively. O3_dis.mat, S4_dis.mat, X11_dis.mat are the continuous discriminant distance output (magnitude output). 3. Results The mutual information plots on the training data set are as shown in Fig1, Fig2, Fig3. Fig 1. The mutual information of subject O3 Fig 2. The mutual information of Subject X11 Fig 3. The mutual information of Subject S4