| Abstract [eng] |
Electroencephalography (EEG) is a non-invasive electrophysiological technique to record and monitor the electrical activity of the brain from electrodes placed along the scalp. In last two decades, a wide range applications of EEG signals have been proposed such as, not limit to, epilepsy detection and prediction, diagnosis of sleep disorders, depth of anesthesia and coma, Brain Computer Interfaces (BCI) applications, etc. The utilization of recorded EEG signals might be hindered due to artifacts and noise of physiological and non-physiological sources, e.g., eye blinks (EB), eye movements, muscle contractions, cardiac activity, electrical shift and linear trend (ESLT), power line interference, etc. The influence of these artifacts might lead to the wrong statistical and physiological analysis of the brain activity. In particular for long-term EEG recordings or autonomous BCI systems, where the noise contribution is random and cannot be supervised by the human expert, the automatic elimination of EEG artifacts is a necessary step before further processing and analysis. Amongst artifacts with the physiological origin, EB, due to its large amplitude and inevitable frequent appearance, is considered to have the most adverse influence on the EEG analysis. ESLT artifacts may emerge in EEG signals due to the temporary shift or lose of electrodes during the recording. In this thesis, low complexity approaches based on the Stationary Wavelet Transform (SWT) for the elimination of EB and ESLT artifacts from EEG signals are proposed. The main novelty of this research is to employ skewness and kurtosis to stop the decomposition level of SWT once it reaches the artifact components. It is shown that the skewness and kurtosis could be suitable artifact markers for EB and ESLT, respectively. The proposed methods are compared against the Automatic Wavelet Independent Components Analysis (AWICA) and Enhanced AWICA (EAWICA) which were presented for automatic EEG denoising. The performance of all algorithms have been tested on simulated and real contaminated EEG signals. Normalized Root Mean Square Error (NRMSE), Peak Signal-to-Noise Ratio (PSNR), and correlation coefficient (CC) between filtered and pure EEG signals are utilized to quantify artifact removal performance. The proposed approaches show smaller NRMSE and larger PSNR, and CC values compared to the other methods. Furthermore, the speed of execution for the proposed algorithms are considerably shorter than the algorithms for comparison, which makes them more suitable for the real-time processing. |
