EEG-based eight class, single trial classification of imaginary wrist movements

In this paper a two-stage, eight class classifier for EEG-based brain computer interfaces (BCIs) is proposed. The first stage classifies left vs. right imaginary wrist movement and the second stage classifies four different movements on each of the wrists. Only results for the second stage are shown here. The average classification rate for five subjects was 80.2plusmn9.4% for the right wrist and 80.0plusmn7.8% for the left one. Classification was performed with a recurrent Elman neural network using complex and real value features in the time-frequency domain as input. The latter were obtained using the Gabor transform. Over 70% of the useful features were found to belong to the gamma EEG band.