OBJECTIVE: To classify the EEG data recorded in mu rhythm-based cursor control experiments with 4 possible choices. METHODS: The algorithm included preprocessing, feature extraction, and classification. Two spatial filters, common average reference and common spatial subspace decomposition, were used in preprocessing to improve the signal-to-noise ratio, and then two features were extracted based on the power spectrum and the time course of the mu rhythm respectively. A Fisher ratio was defined to select channels in feature extraction. A 2-dimensional linear classifier was trained for final classification. RESULTS: Two types of classifiers were trained for the training dataset. The uniform classifier gave a classification accuracy of 76.4%, and the classifier trained by leave-one-out method gave a classification accuracy of 74.4%, both higher than the online accuracy 69.5%. The uniform classifier was applied to the test dataset and the classification accuracy was 65.9%, lower than the online accuracy 73.2%. CONCLUSIONS: Spatial filtering can give a notable improvement in classification accuracy. The time course of the mu rhythm, as well as the power of the mu rhythm, shows difference between the 4 targets, and can contribute to the classification. SIGNIFICANCE: The spatial filtering, feature extraction and channel selection methods in the algorithm will provide some practical suggestions for further study on the mu rhythm-based brain-computer interface.
OBJECTIVE: To classify the EEG data recorded in mu rhythm-based cursor control experiments with 4 possible choices. METHODS: The algorithm included preprocessing, feature extraction, and classification. Two spatial filters, common average reference and common spatial subspace decomposition, were used in preprocessing to improve the signal-to-noise ratio, and then two features were extracted based on the power spectrum and the time course of the mu rhythm respectively. A Fisher ratio was defined to select channels in feature extraction. A 2-dimensional linear classifier was trained for final classification. RESULTS: Two types of classifiers were trained for the training dataset. The uniform classifier gave a classification accuracy of 76.4%, and the classifier trained by leave-one-out method gave a classification accuracy of 74.4%, both higher than the online accuracy 69.5%. The uniform classifier was applied to the test dataset and the classification accuracy was 65.9%, lower than the online accuracy 73.2%. CONCLUSIONS: Spatial filtering can give a notable improvement in classification accuracy. The time course of the mu rhythm, as well as the power of the mu rhythm, shows difference between the 4 targets, and can contribute to the classification. SIGNIFICANCE: The spatial filtering, feature extraction and channel selection methods in the algorithm will provide some practical suggestions for further study on the mu rhythm-based brain-computer interface.