Information Technology Reference
In-Depth Information
that the ERS engendered by this source is in this relationship: ERS(UE) > ERS
(EE) > ERS(EC), with both inequality signs indicating a signi
cant difference
(
p
< 0.05).
8.9.6 Classification of Single Trials
The Ne source alone leads to better accuracy in classifying error trials as compared
to the theta source alone (
p
< 0.01). The theta source leads to better accuracy for
classifying correct trials (
p
= 0.028). These corroborate the conclusion that the ERP
and ERS represent different phenomena of the ErrP. When looking at the average
classi
cation rate (Te + Tc)/2, with Te the classi
cation rate of error trials and Tc
the classi
cation rate of correct trials, one see that the use of both components leads
to better results for 14 subjects out of 19. The use of both components increases the
mean classi
cation rate on the 19 subjects from 67 % up to 71 %. We performed a
repeated-measure two-way ANOVA with factor
(error vs. correct) and
“
feature
”
(Ne source ERP, theta source ERS, both). It revealed a main effect on the
“
“
type
”
type
”
factor (
p
< 0.01) with correct trials being better classi
ed than error trials and
a
interaction (p = 0.013), demonstrating that the use of both the
ERP feature and the ERS feature in the source space improves the performance of
single-trial classi
“
type
”
x
“
feature
”
cation. It should be noticed that with a total of 72 trials per
subject, training set included only a mean of 17 single trials for the error condition;
thus, the classi
cation task for this data set is hard since the training sets include
very few examples of error trials.
Knowing that the error expectation has an in
uence on the theta ERS, we have
looked at classi
cation results for expected and unexpected errors, for the theta ERS
components, and the Ne ERP components. Classi
cation performance is higher for
unexpected errors (mean Te = 62 %) than for expected errors (mean Te = 47 %)
(
p
= 0.011) when using the theta ERS component. On the other hand, results are
equivalent (mean Te = 63 % for unexpected errors and mean Te = 64 % for
expected errors) using the Ne ERP component for classi
cation
using the theta ERS component performs poorly on error trials only for expected
errors. As a consequence, in the case of a system where errors are unexpected, the
classi
cation. Thus, classi
cation using the theta ERS component as compared to using the Ne ERP
component would allow similar results for error trials and better results for corrects
trials, leading to a better average classi
cation accuracy.
8.10
Conclusions and Discussion
We have described a blind source separation approach requiring only estimation of
second-order statistics of data, that is, covariance matrices. The method, which is
well grounded on current theory of volume conduction, is consistently solved by
means of approximate joint diagonalization of a set of covariance matrices, whence
