Biomedical Engineering Reference
In-Depth Information
rently active processes such as the retraction and rotation of the eye bulb [19] as well
as the closure and reopening of the eyelids [20].
Although the utility and accuracy of regression-based ocular correction is a mat-
ter of ongoing discussion [21], it is widely used, and a number of implementations
and variations exist. The most commonly used are those by Gratton et al. [22] and
Semlitsch [23]. Both algorithms share the general mechanism of first finding
blink-afflicted data stretches in the EOG channel(s) with (different) thresholding
techniques. Based on the blink stretches found, both algorithms then calculate the
regression of the eye channel(s) with each individual EEG data channel, and correct
the EEG data with EEG
′=
EEG
− β ×
EOG, where
β
is the regression weight for a
given channel.
The Gratton et al. algorithm [22] has two further characteristics that are worth
noting: First, the raw averages for each condition are subtracted from each data seg-
ment prior to regression calculation and are added back in before the correction is
performed. Second, the algorithm calculates and applies separate regression coeffi-
cients for data time ranges inside of blink stretches and for those outside of blink
stretches, which can easily lead to the creation of step discontinuities during regres-
sion, since time ranges corrected with (slightly) different regression coefficients bor-
der directly onto one another. The raw average subtraction is done under the
assumption that the measured data can be expressed as EEG
NOISE
and that all four components are uncorrelated. Under the assumption that EEG
tends towards 0
+
ERP
+
EOG
+
V with averaging and after subtraction of the raw average ERP,
the regression would be calculated on the EOG
μ
NOISE components alone, which
is, of course, desirable. However, this also implicitly assumes that the EOG compo-
nent is temporally stochastic (not stimulus contingent), which clearly is not the case
for many paradigms used today (e.g., visual search, emotion induction paradigms).
Thus, a varying amount of stimulus-evoked EOG activity is subtracted along with
the raw ERP average, and the regression is then based on the residual, non-stimu-
lus-contingent portion of the EOG activity alone.
A more general problem of regression-based artifact correction is that this
approach assumes stability of the artifact over time, which is not always given, espe-
cially with experimental paradigms that are monotonous and fatigue inducing.
Another problem results from the fact that the regression-based correction only
works properly if the eye electrodes are placed completely perpendicular to one
another. If this is not the case, then the eye artifact data is not completely linearly
represented in the artifact time stretches of the EEG channels, which can result in
over- or undercorrection. However, the most grave problem inherent to this method
is that it assumes a directional relationship between EOG and EEG where it is
assumed that the EOG activity alone is causing the commonality found in the “arti-
facts” in the EEG; in reality, however, the EEG activity present is as likely to influ-
ence the EOG channel recordings. Correcting the EEG readings based on the
regression weights may therefore remove substantial amounts of desired EEG
(effect) activity along with the true eye movement-borne artifacts. For these reasons
and from our own experience, independent component analysis is clearly favored
over regression-based approaches for the correction of eye blinks and other
artifacts. This approach is discussed in detail in the following section.
+
Search WWH ::
Custom Search