Feature combination for classifying single-trial ECoG during motor imagery of different sessions

The input signals of brain-computer interfaces (BCIs) may be either scalp electroencephalogram (EEG) or electrocorticogram (ECoG) recorded from subdural electrodes. To make BCIs practical, the classifiers for discriminating different brain states must have the ability of session-to-session transfer. This paper proposes an algorithm for classifying single-trial ECoG during motor imagery of different sessions. Three features, derived from two physiological phenomena, movement-related potentials (MRP) and event-related desynchronization (ERD), and extracted by common spatial subspace decomposition (CSSD) and waveform mean, are combined to perform classification tasks. The specific signal processing methods utilized are described in detail. The algorithm was successfully applied to Data Set I of BCI Competition III, and achieved a classification accuracy of 91% on test set.