Reduction of irrelevant and redundant data from TFRs for EEG signal classification

Time-frequency representations (TFR) are one of the most popular characterization methods for non-stationary biosignals. Despite of their potential advantages, these representations suffer of large quantity of redundant and irrelevant data which makes them difficult to use for classification purposes. In this work, a methodology for reduction of irrelevant and redundant data is explored. This approach consists on removing irrelevant data, applying a relevance measure on the t-f plane that measures the dependence of each t-f point with the class labels. Then, principal component analysis (PCA) and partial least squares (PLS) are used as non-supervised and supervised linear decomposition approaches to reduce redundancy of remaining t-f points. Results show that the proposed methodology improves the performance of classifier up to 3% when no relevance and redundancy on TFRs is reduced.