Abstract
To enhance the reliability of motor imagery based brain-computer interface, we present a study that considers subject-based optimization of feature extraction and classification. In particular, wavelet-based feature extraction performed on different bands was optimized over available selections of wavelet family, length and number of decomposition levels. Likewise, the classification step considers three general families of classifiers whose parameters are optimized in a similar manner. Such optimization was performed for each subject whereby processing parameters are selected based on the best performance obtained in the training session. We report the results obtained from applying this approach to the BCI competition 2008 dataset 2b (Graz) and demonstrate that such optimization provides results that outperform previous methods.