P.E. Volynsky, E.L. Masherov, G.A. Shekut-ev, Yu.V. Obukhov

EEG is widely used to studying brain activity. Digitizing of EEG recordings using modern equipment allows one to utilize a wide range of mathematical algorithms of digital signal processing. However, most of current digital analysis algorithms are sensitive to distortions of the original signal by non-physiological artifacts and noise. A variety of different methods have been already proposed to solve this problem. However these either require direct involvement of the researcher or based on additional measurements, or cannot remove artifact activity to a sufficient extent. Here we propose a novel analysis method which is based on independence of the artifact signals in different EEG channels and on the representation of electrical activity in one channel through activity in other channel (cross-regression, CR). This method does not use any a priory information and thus is fully automated. Testing of the proposed method using model signals and real EEG recordings has demonstrated its effectiveness for elimination of the non-correlated EEG components such as electromyogram or movement artifacts. Application of the CR analysis may lead to noise amplification in EEG channels that contain high amplitude long lasting artifact signals, especially when analysing short EEG recordings. Pre-processing of the EEG data using frequency filtration or amplitude thresholding normally allows eliminating such noise amplification. However this approach requires choosing a priory settings of threshold parameters for frequency or amplitude values. In order to remove such pre-processing step we tested a CR based Least Squares algorithm and a Least Amplitude Module algorithm. We demonstrate that the Least Amplitude Module algorithm allows artefact elimination without significant distortion of the source EEG signal. Thus we conclude that the Least Amplitude Module algorithm can be directly applied for artefact elimination without the pre-processing step.