Information Technology Reference
In-Depth Information
jointly diagonalizes a total of 18 matrices. For solving the approximate joint
diagonalization, we employ the iterative algorithm proposed by Tichavsky and
Yeredor (
2009
), which is fast and in our long-lasting practice has proven robust.
Once estimated the 31 sources, they were inspected analyzing their ERP, ERS,
topographies, and the mutual information criterion between the source and the error
class (Grosse-Wentrup and Buss
2008
). Meaningful sources were localized in a
standard brain using the sLORETA inverse solution (Pascual-Marqui
2002
)as
implemented in the LORETA-Key software. This software makes use of revisited
realistic electrode coordinates (Jurcak et al.
2007
) and the head model (and cor-
responding lead-
eld matrix) produced by Fuchs et al. (
2002
), applying the
boundary element method on the MNI-152 (Montreal neurological institute,
Canada) template of Mazziotta et al. (
2001
). The sLORETA-key anatomical
template divides and labels the neocortical (including hippocampus and anterior
cingulate cortex) MNI-152 volume in 6,239 voxels of dimension 5 mm
3
, based on
probabilities returned by the Demon Atlas (Lancaster et al.
2000
). The coregis-
tration makes use of the correct translation from the MNI-152 space into the
Talairach and Tournoux (
1988
) space (Brett et al.
2002
). Source localization was
conducted on each participant separately, normalized to unit global current density
(the input of the inverse solution is a vector estimated by BSS up to a scale
indeterminacy) and summed up over participants in the brain space.
8.8.7 Classification of Single Trials
For classifying single trials, data were band-pass-
ltered using an order four But-
terworth FIR
filter with linear phase response between 1 and 10 Hz for the ERP
component and 4
-
8 Hz for the ERS component. Data were then spatially
filtered
using the results of the BSS analysis. Only samples corresponding to 250
750 ms
were kept. For the ERP component, we used the temporal signal down-sampled at
32 Hz, providing 16 samples (features) for the classi
-
cation. For the ERS component,
we used the square of the temporal signal (power) down-sampled at 32 Hz, providing
16 samples (features) for the classi
cation as well. This procedure assigns to each
component equal chance for classi
er, we employed a linear
discriminant analysis (LDA). One hundred random cross-validations were performed
with the classi
cation. As a classi
er trained on a randomly selected set containing 80 % of the data (both
errors and corrects) and then tested on the remaining data.
8.9
Results
8.9.1 Behavioral Results
All subjects performed the task with a convenient error rate, with mean (sd) = 22.2
(4) % and a quasi-equal repartition of expected and unexpected errors, with mean
(sd) = 10.4 (4.3) % and 11.8(3) %, respectively. Reaction time was higher for error
