Improvement of Single-Trial EEG Classifier Accuracy Based on Combination of Optimal Spatial Filters and Time-Domain Features

Common spatial pattern (CSP) is a popular technique in feature extraction for brain-computer interface (BCI). However, CSP algorithm itself does not perform well since the estimation of covariance matrices is quite sensitive to training data. This causes over fitting in some cases, especially when the training set is small. Moreover, this method may result in poor outcomes because it just computes features on spatial domain but omit those on other domains. In this paper, we propose a simple yet effective approach. Through improving the CSP algorithm, optimal spatial filters with highest discriminative ability will be extracted. Concurrently, we also incorporate some time-domain information into feature vectors to make the signal presentation become more sufficient. Experiment results show that this is a promising method for an electroencephalography (EEG) - based brain-computer interface system.