EEG artifact removal for ambulatory epileptic seizure prediction applications -

Abstract

Electroencephalography (EEG) is considered a primary tool for monitoring the electrical activity of the brain. Recent advances in wearable sensing techniques allow continuous and mobile monitoring of EEG signals during daily life activities. However, such method of EEG recording is prone to different sources of artifacts: eye-movement, electrode movement, muscle contraction, line noise, head movement and others. Among these sources, motion-related artifacts are a major challenge for clean EEG data acquisition. The significant effect of motion artifacts is evident in two main aspects. First, they overlap with all EEG frequency bands. Second, they spread over the entire scalp affecting all sensing electrodes. For some neuro applications, such as epileptic seizure detection and prediction, high quality EEG signals are required to accurately depict the electrical activity of the brain and thus track seizure markers for correct classification. The main focus of this thesis work is to record EEG data with defined motion artifacts in a controlled lab environment, to utilize various algorithmic methods for the effective elimination of motion artifacts, to assess the performance of the adopted artifact removal technique using statistical measures, and to employ the latter technique in the application of seizure detection and prediction. The adopted approach for artifact removal is based on applying independent component analysis (ICA) as a blind source separation technique for removing mobility artifacts from EEG data. The quality of the reconstructed EEG signals is assessed first using various statistical measures and then through investigating the seizure-prediction and seizure-detection capabilities of the reconstructed signals as opposed to the capabilities of the original noise-free signals. For detection and prediction purposes, the EEG signals are analyzed by extracting distinctive features using an N-gram based algorithm. These features are used to train a predictive model, which in turn used it to classify EEG s