Quantitative evaluation of automatic ocular removal from simulated EEG signals: regression vs. second order statistics methods.

Analysis of the EEG by means of spectral parameters permit to evaluate the influence of a drug and to diagnose dysfunctional states in neurology, psychiatry and psychopharmacology. Eye movement artifacts contaminate EEG signals and can produce errors in this analysis. Regression based technique is considered the 'gold standard' artifact removal procedure and other techniques have been developed the last years, but few works have shown an objectively evaluation of the efficiency of these methods because it is impossible to record pure EEG and EOG signals. In this study, an artificially reproduction of bidirectional contaminated EEG and EOG data is proposed in order to simulate a real case. A comparative study between automatic second-order statistics techniques (PCA, AMUSE and SOBI) and multiple regression analysis is performed. Effectiveness of removal techniques is evaluated by calculating the errors in spectral parameters between sources and corrected EEG signals. Average values and topographic brain distribution of these errors are considered. Errors are located in the anterior leads especially in the frontopolar ones. Results show that AMUSE and SOBI methods preserve more cerebral activity than other techniques. We conclude that AMUSE and SOBI algorithms overcome the limitations of the regression based approach in the bidirectional contamination between ocular and neural activity.