Digital Signal Processing and System Theory

Talk Hui Xu

Automatic Identification and Suppression of Artifacts in EEG data using ICA and Statistical Thresholding

Date: 19.09.2011, 13:00 h - 14:00 h, Room: Aquarium

Hui Xu
CAU, Kiel, Germany


Electroencephalography (EEG) is a medical technique used to measure the brain function of human beings by analyzing the scalp electrical activity generated by the brain structure. Usually the EEG data are typically contaminated with interference waveforms, called artifacts, which could be physiologic (e.g., eye blinks, heart beating, muscle activities) and extraphysiologic (electrode popping, 50Hz line noise power). Many methods have been proposed to remove the artifacts from EEG recordings. There exist the manual methods, which require the presence of an experienced doctor or clinician to detect, by visual inspection, the artifactual components or epochs and then remove them to correct the data. However, it represents a very time-consumption problem due to the vast amount of data. Moreover, the efficiency of manual methods is poor, since some artifacts might be left while some useful components could be deleted by mistake. Therefore, efficient and robust automatic methods for suppression of the artifacts are strongly required. From the last decade there has been a great development of the methods proposed to remove artifacts automatically from EEG recordings. However, most of them focus and remove only some of them, namely the eye movements. Hence, one main difficulty in the automatic EEG analysis is the detection of the different kind of artifacts added to the EEG signals during the recording sessions.
In this Thesis a new identification procedure based on an efficient combination of Independent Component Analysis (ICA), Wiener Filter, and statistical threshold values of parameters for fully automatic artifact suppression was presented. This proposed approach has been called Automatic Identification and Suppression of artifacts based on Fast Independent Component Analysis (FastICA) and Thresholding (AISFT). FastICA was then first applied to decompose the original EEG signals into independent sources. Parameters such as Hurst exponent, kurtosis, correlation coefficient, variance and amplitude range were used to identify the artifactual components and hence suppress the intervals where the artifacts were. In this way, the artifacts were suppressed but the desired information still remained, compared with the current manual methods where the valuable information could be lost when removing either the distorted epoch or the complete component. The AISFT method was applied to both, healthy and epileptic EEGs. Implementation results were compared and analyzed in time and frequency domains. Finally, feasible plans for future work were given based on the existing left problems.