Temporal Window based Feature Extraction Technique for Motor-Imagery EEG Signal Classification

Electroencephalogram (EEG) based motor-imagery classification is one of the most popular Brain Computer Interface (BCI) research areas due to its portability and low cost. In this paper, we have compared Wavelet Energy-entropy based different prediction models and empirically proven that temporal window based approach in motor-imagery classification provides more consistent and better results than popular filter-bank approach. In order to examine the robustness and stability of the proposed method, we have also employed multiple types of classifiers at the end and found that mix-bagging (bagging ensemble learning with multiple types of learners) technique out-smarts other frequently used classifiers. In our study, BCI Competition II Data-set III has been used with four experimental setup: (a) The whole signal (for each trial) as one segment, (b) The whole signal (for each trial) is divided into non-overlapping segments, (c) The whole signal (for each trial) is divided into overlapping segments, and (d) The filter-bank approach where the whole signal (each trial) is segmented based on different frequency bands. The result obtained from the experiment (c) i.e. 91.43% classification accuracy which outperforms the other approaches not only in this paper but to best of our knowledge it is the highest performance for this dataset so far.