Autonomous noise removal from EEG signals using independent component analysis

Electroencephalography (EEG) based brain-computer interface (BCI) systems have drawn wide interest from researchers in biomedical sector due to their low cost and non-invasive features. Such interfaces have become viable to control systems utilizing user created neural impulses. Nevertheless, EEG signals are easily susceptible to noise contamination from ocular and muscular movements. Extensive contamination within the signal could lead to mis-classifications and compromise user's safety. The independent component analysis (ICA) algorithm has been used to successfully extract noise contamination. However, it requires prior knowledge and experience regarding visual characteristics of the noise signals to effectively extract them. In this paper, a novel autonomous noise extraction approach is presented that uses both constant and variable threshold parameters to identify and extract noise features from EEG signals acquired using ICA. An artificial neural network (ANN) was used to validate the effectiveness of filtering techniques. The EEG Motor Movement/Imagery dataset, available on PhysioNet, was used to in this study. Six imagery tasks were divided into two sets for 30 subjects and classified using the ANN. An average classification accuracy of 92.11% and 91.99% was achieved for Set 1 and Set 2 using the proposed autonomous extraction techniques. Overall, the accuracies were improved by an average of 0.81% and 0.75% among both sets for all the subjects.