Jorge del Río Vera, BCI Competition 2003, data set III Classification of imagery left or right hand movements 1.- Algorithm The algorithm used to classificate the signals is based in a MLP neural network trained by the backpropagation method. Inputs patterns are obtained from the application of a PCA-based algorithm (Principal Components Analysis). The training set was analized to get an inicial view of the important features of the signals. An ERD diagram was calculated for each class (class 1 or class 2) separately (Figure 1). Figure 1.- Signals are filtered to obtain the Mu rythm which is associated with movements [1]. ERD diagram of each filtered class set is computed in order to obtain the main features of each channel. [see DOC file] Once the time interval, where the desynchronization take place, is clearly shown (it is set between the seconds 4 and 5), the principal component analysis is computed for each class and each channel, only the interval between 4 a 5 is taken into account, and only the five principal components for each class and channel are used. This analysis results in 30 different patterns of 128 samples. After that, signals are divided in one second windows (only the time period between 4-9 seconds is used ) overlapped 127 samples from the previous (to achieve one classification for each sample). The windows obtained from the test signals are correlated2 with the ones obtained from the PCA analysis and only the maximum values are used (the maximum value shows the similarity between the mode and the signal)3. This proceeding results in five values to each channel and each class. Each set of five values are condensed in one by Eq.1 (Eq 1shows the proceeding for class 1 the proceedings for class 2 is the same). [Eq. 1] The result is a six component vector for each signal window, 1 for each channel and each class. These vector are used to train a MLP backpropagation network with an inner layer with ten neurons, and the output layer with 24. This neural network used to classify to the processed test signals. 2.- Results Results are given in a matlab file called "results.mat", it contains a variable called "results" with the following structure: Results 640x140 column one have the results for the first set of three signals of the x_test vector. Classification coeficients are given in a -10 to 10 range according to the requirements, and only for the time period between 4 to 9. 1 - J.R. Wolpaw, "An EEG-based brain-computer interface for cursor control", Electroenceph. Clin. Neurophysiol, Vol. 78, pp. 252-259, 1991. 2 Test signal from C3 is correlated with PCA-C3 mode 1 to 5 for class 1 and class 2, test signal from Cz is correlated with PCA-Cz mode 1 to 5 for class 1 and class 2 and test signal from C4 is correlated with PCA-C4 mode 1 to 5 for class 1 and class 2. Then the maximun value of each correlation is selected. 3 Notation: M1_2C3 --> Maximum value of the correlation between mode 1 of PCA analysis for class 2 and C3 with C3. 4 - C. Anderson, "Classification of EEG signals from four subjects during five mental task", In Proc. Of the conf. On engineering applications in neural networks (EANN´96),pp. 407-414, Systems Engineering Association, 1996.