Time-Frequency Features Combination to Improve Single-Trial EEG Classification

摘要

In this paper, we propose a combination of two simple feature extraction methods from time and frequency domain to improve singe-trial EEG classification in self-paced BCI. 'Bereitschaftspotential' (BP) features from time domain and event-related desynchronization (ERD) features from frequency domain are merged and feed into four different classifiers which are probabilistic neural network (PNN), support-vector machine (SVM), K-nearest neighbor (KNN), and Parzen classifier (PC). Results using BCI competition 2003 |1| dataset IV are showing that the combined features are quite discriminative as we reached an accuracy on the test set ranging from 82% to 85% whereas the winner of the competition on this data set reached 84% using three types of features |2,3|.