Abstracts

The objective of this study was to devise an automated system to remove artefacts from ictal scalp EEG, using independent component analysis (ICA). Current ICA methods rely on a tedious visual identification of artefactual sources among extracted components, hence the desire for a fully automated method. 69 ictal scalp recordings from 16 epileptic patients were analyzed. The seizures were contaminated by various artefacts such as eye movements, EMG, and patient movement. For each seizure, ICA was applied to a 30s interval beginning approximately 10s before the suspected seizure onset. The extracted independent components were divided into 2s epochs and each epoch was labelled as either EEG or non-EEG. Data from half of the patients was used as a training set to induce a Bayesian network classifier, while the remaining data was reserved for use as a validation set. The classifier used the following features: component negentropy, epoch variance, spectral entropy between 5 and 30Hz, and relative power in several frequency bands. The spatial distribution of each component was then fitted with a dipole, the position of which was also used as a feature whenever the residual variance of the fit was less than 20%. For each 2s epoch, the output of the classifier was the probability that the epoch represented EEG activity. A component was considered to be artefactual if the sum of the probabilities for its 15 epochs was less than 4. To evaluate the performance of the system, an expert neurologist reviewed the original seizure recordings in the validation set and compared them with the records reconstructed after rejection of the artefactual components. A qualitative score was given with respect to the reduction of artefactual activity and the preservation of EEG activity. The validation set contained 33 seizures from 8 patients. The system correctly classified 5047/5920 (85%) EEG epochs and 5628/6950 (81%) non-EEG epochs. The reviewer noted that 23 seizures were contaminated by a significant amount of artefacts. After the automated rejection of artefactual components, the reviewer determined that the majority of artefacts were removed in 17 seizures, while 5 of them had minor improvements, and only 1 record had no reduction of artefactual activity. The system preserved all the EEG activity in 17 seizures, attenuated some minor EEG activity in 4 cases, and removed some significant ictal activity in the remaining 2 cases. The automated method was also applied to the 10 seizures that did not contain significant artefacts; in 9 cases, all the EEG activity was preserved, while the remaining case had only minor EEG attenuation. Temporal and spatial features of ICA components were used in a Bayesian framework to classify component epochs as either EEG or non-EEG. This allowed the proposed system to automatically eliminate a large proportion of artefactual components in ictal scalp recordings, while minimally affecting the EEG activity. (Supported by NSERC CGS-M, CIHR MOP-10189.)