Bahareh Ahkami - Department of Biomedical Engineering-Bioelectric at Amirkabir University of Technology (AUT)
Farnaz Gassemi - Department of Biomedical Engineering-Bioelectric at Amirkabir University of Technology (AUT)
Ali Nakisai - Department of Biomedical Engineering-Bioelectric at Amirkabir University of Technology (AUT)

BackgroundElectrooculogram (EOG) is one of the most important artefacts in Electroencephalogram (EEG) signals. These artefacts are barriers between EEG signals and useable cognitive information that can be inferred from brain signals. Various methods have been introduced to reduce these artefacts in the preprocessing stage of signal processing. This step is crucial for making further decisions based on clean signals. Removing artefacts shouldbe accurate since signals with brain information should not be lost; Moreover, todays technology is heading towards online tasks and fast processing is an important factor especially in Brain Computer Interface (BCI) systems. In this work, we are focusing on EOG, as it is a common artefact and can be observed in most EEG signals. Independent Component Analysis (ICA) and its combination with other methods are common ways to reduce EEG artefacts by means of finding and eliminating the EOG components. 2. MethodIn this paper modified ICA-RLS was used to reduce EOG artefact from EEG signals and the results were compared to the results of ICA-Correlation method which is a common way to remove EOG artefacts. In ICA-Correlation the correlations between EOG channel and all the independent components are calculated, nd the components with the correlations higher than a predefined threshold will be removed. Then the remaining components are back projected to channel domain. In contrast, in ICA-RLS method, each independent component is filtered out with respect to EOG channel. As the magnitude of output signal of the adaptive filter gets closer to zero, it is concluded that the component is more similar to EOG channel. As a modification, to overcome the possible effect of variations in magnitudes of every component on the resulting output of RLS filter, a normalization step on the RLS filter output magnitude is introduced. The resulting value is called the Comparison Coefficient (CC). EEG signals of 12 healthy subjects, acquired with a 19 electrode EEG system was used in this research. 3. ResultsEvaluations showed that CC is a small value for the thought-to-be EOG component and for other components, it is 5 to 20 times larger. In ICA-Correlation, the same comparison coefficient was utilized; however, the same diversity in the CC among different components as in ICA-RLS was not eventuated, so it may not be possible to firmly reject one component. This evaluation is performed once on the entire signal, and once on short,randomly-selected pieces of the entire signal in order to demonstrate the efficiency of these two methods in online scenarios.4. Conclusions Results show that ICA-Correlation is not a trustworthy method especially when analyzing short-piece signals, and in some points it leads to a wrong selection of the EOG component, while ICA-RLS proves to be an efficient method in all cases. Based on the results, it was shown that the modified normalized ICA-RLS method doesbetter than the classic ICA-RLS one. Furthermore, performance evaluations show that the proposed modifiedICA-RLS works 40 percent faster than the ICA-Correlation method.