Electrical signals detected along the scalp by an Electroencephalogram (EEG), but that originate from non-cerebral origin are called artifacts. Especially when these artifacts are produced by the human body we talk about biological artifacts. The most common biological artifacts are the electrical signals produced by ocular and heart activity. EEG data is almost always contaminated by such artifacts. The last decade Independent Component Analysis (ICA) has a crucial role in neuroscience and it takes great attention for artifact rejection purposes. According to ICA’s methodology, EEG signals are decomposed to statistical Independent Components (IC) and then an EEG specialist is called to recognize the artifactual ICs. Some of the major limitations of the current approach are that the aforementioned selection is subjective, it demands a high skill EEG operator, it is time consuming and it cannot be applied in online processing. Our study employs machine learning techniques in order to recognize the contaminated ICs with ocular or heart artifacts. More specific 19-channel EEG datasets from 86 normal subjects were decomposed using ICA (19×86=1634 ICs in total). Then three independent observers marked an IC as artifactual if it includes ocular or heart artifacts, otherwise it was marked as normal. Then kurtosis was computed in short segments with 1250 sample points fixed length without overlap for each IC. The mean kurtosis value was computed for each IC and the Naïve Bayes Classifier (NBC) classifier was adopted in order to classify the ICs as artifactual or normal. The results suggest that the NBC has correctly classified 1611/1634 ICs (98.5924 %) so it can be suggested that kurtosis seems to be convenient for the classification of contaminated ICs by ocular or heart artifacts.
|Title of host publication
|XII Mediterranean Conference on Medical and Biological Engineering and Computing 2010
|Number of pages
|Published - 2010
- naïve bayes classifier