Three-class classification of motor imagery EEG data including 'rest state' using filter-bank multi-class Common Spatial pattern

Our purpose is to develop the 3-class Brain Machine Interface (BMI) incorporating the classification of resting state using Electroencephalography (EEG). Conventionally the most of BMI systems only accept EEG data when a subject performs some kind of task, such as motor imagery and gaze at visual stimuli. However, performing task causes fatigue of the subject. It is therefore important to develop classification algorithm for BMI system that utilizes rest state-EEG as one of the classes. The 3 classes we defined in this experiment were: (1) motor imagery of moving right hand; (2) motor imagery of moving left hand; and (3) rest state. And, we also measured EEG in an actual moving task (finger tapping) to ascertain validity of algorithm. We extracted feature vector using Finite Impulse Response (FIR) digital filter Filter Bank and multi-class Common Spatial Filter (mCSP) from EEG data, selected the feature by Mutual Information (MI), and made three 3-class classifiers using Random Forest (RF). The mean classification rate was 56.7±4.43% at motor imagery task and 88.7±4.54% at actual finger tapping task. And we compared the time required to extract features and compute classifiers with those of other methods. Our method is effective to some extent. (1) parameter selection time was better than choosing single band-pass filter that best discriminate classes among possible options of frequency bands; and (2) accuracy rate was better than our previous method using majority vote.