Handling Artifacts in BESA

Module information
Modules	BESA Research Basic or higher
Version	5.2 or higher

Artifact correction always aims at extracting unwanted signals like EOG, EKG or external noise from the data, while leaving all brain activity of interest as undisturbed as possible. To achieve this, artifact and brain topographies must be separated. Depending on whether one is dealing with spontaneous or evoked activity, different approaches for artifact correction are appropriate. However, if the data contain only a few artifacts and a sufficient number of artifact-free trials are retained, using artifact rejection rather than correction is advised as this avoids the distortion resulting from correction.

Principles of artifact correction

For artifact correction, artifact and brain activities must be identified and separated. In general, artifact and brain topographies will be spatially correlated. Hence, a simple regression or the projection of the data onto the subspace orthogonal to the artifact topographies will severely distort the data. For a correction without distortion, it is not sufficient to define the artifact topographies (to be removed), but it is equally necessary to create a model or a spatial description of the brain topographies (to be retained).

Artifact correction Methods

Adaptive artifact correction

This method estimates the brain activity from the data currently displayed on the screen. The data is scanned in specified time intervals. Those segments are considered to represent brain activity where 1) the correlation between data and artifact topography does not exceed a certain threshold and 2) the signal amplitudes are below a specified threshold. Of the remaining segments a principal component analysis (PCA) is performed. All PCA components explaining more than the minimum variance specified in the box Adaptive Model: PCA Topography are maintained. They span the brain signal subspace.

In a next step, the recorded data is decomposed using all topographies into a linear combination of brain and artifact activities. Thus, the estimated artifact signals are much less overlapped with brain activity and can be subtracted from the original signals without much distortion. This approach is recommended, in particular, for the review of continuous EEG or MEG data.

Adaptive artifact correction

Here, brain activity is modeled by a model consisting of multiple equivalent current dipoles. The artifact topographies are added to this model and the combined model is then applied to the recorded data. Again, the estimated inverse signals separate the brain activity associated with the surrogate sources from the artifacts to a high degree. Thus, the artifact signals can be subtracted without considerable distortion of the activities originating in the modeled regions. This approach considers the activity in the modeled brain regions while the on-going EEG is not modeled accurately.

Therefore, the surrogate method is especially recommended for the correction of data to be averaged if the average signal is smaller than the EEG or MEG background. In this case a model cannot be estimated from the on-going data. Therefore, a-priori knowledge of the involved brain regions should be employed to create an appropriate surrogate model.

Subspace projection (SSP, regression)

This approach has been commonly applied in the literature. SSP does not contrast artifacts and brain activity. Rather, the complete subspace spanned by the artifact topographies is projected away from the recorded data. This leads to un-distorted data only in the highly unlikely case when artifact and brain activity have exactly orthogonal topographies. This is generally not the case in real data. In the likely event that evoked brain activity has a topography correlated with the artifact, this method removes the correlated fraction of the brain activity. As one of the negative consequences, maps of the corrected brain activity will be severely distorted after SSP correction.