<![CDATA[Motor imagery (MI) responses extracted from the brain in the form of EEG signals have been widely utilized for intention detection in brain computer interface (BCI) systems. However, due to the non-linearity and the non-stationarity of EEG signals, BCI systems suffer from low MI prediction rate with both known and unknown influncing factors. This paper investigates the impact of visual stimulus, feature dimensions and artifacts on MI task detection rate, towards improving MI prediction rate. Three EEG datasets were utilized to facilitate the investigation. Three filters (band-pass, notch and common average reference) and the independent component analysis (ICA) were applied on each datasets, to eliminate the impact of artifact. Three sets of features where extracted from artifact free ICA components, from which more relevant features were selected. Moreover, the selected feature subsets were incorporated into three classifiers, NB, Regression Tree and K-NN to predict four MI and hybrid tasks. K-NN classifier outperformed the other two classifies in each dataset. The highest classification accuracy is obtained in hybrid task EEG dataset. Moreover, accurately predicted EEG classes were applied to a robotic arm control.]]>
