Analysis of EEG Mapping Images to Differentiate Mental Tasks in Brain-Computer Interfaces - Foundations on Natural and Artificial Computation - page 251

4R su s

The classifier has been tested with all three subjects using a cross validation.
As it was mentioned, 3 sessions (75% of the data) have been used to obtain
the models and the remaining session (25% of the data) has been used to test
the classifier. The results obtained are shown in Table 1. Four combinations of
sessions have been obtained showing the success percentage of the classifier. The
success percentage is obtained as the average success of all three tasks. The
success percentage of each task individually is shown for the best user (Subject
1) in Table 2.

Table 1. O ine results - Total success(%)

Subject 1 Subject 2 Subject 3

123-4 68.6 62.8 48.6

124-3 64.6 59.2 44.8

134-2 72.0 60.5 50.4

234-1 51.5 56.4 34.0

Average 64.2 59.7 44.4

Table 2. Oine results for each mental task - Subject 1(%)

Right Word Left

123-4 63.7 67.5 74.9

124-3 55.3 75.9 58.0

134-2 68.5 73.2 75.8

234-1 61.5 30.7 71.7

Average 62.3 61.8 70.1

The success percentages obtained prove that the EEG mapping method for
BCI classification can be an effective new approach in BCI classifiers.
As it has been seen before, a variable of uncertainty has been introduced in
a second group of experimental results. In this case, a threshold value has been
used to reject trials that are not clearly similar to the models (0.8 for Subjects
1 and 3 and 0.75 for Subject 2). It can be verified that the error is reduced
significantly compared to the previous results but affecting the final success
percentage (Table 3). The percentages are obtained as the average percentage
of all three tasks. Figure 6 shows the difference between the averaged results
obtained without uncertainty (first column) and the averaged results obtained
with uncertainty (second column) for each subject.
In Millan (2004) [17], similar results are obtained using a different classifier
with the same data set. With this success percentage and trial time (5 seconds),
the classifier is ready to work in a future online testing.