IEEE Systems, Man and Cybernetics Magazine - January 2022 - 13

others, the accuracy is slightly
low, but they are similar to each
other. Baseline NNs also perform
better than LDA and logistic models,
but the different accuracy
value among all of the seven subjects
is remarkable. The results of
baseline LDA and baseline logistic
regression wildly vary across the
subjects and also overlap for few
cases. The reason of this fluctuation
is originating from the complexity
of the models. NNs are far
complex compared to LDA and
logistic regressions, which is why they do not show a
solid swing in execution.
Table 2 depicts the textual representation of accuracy,
precision, recall, and f1-score for four baseline and TL
methods, where TL shows the highest score. The
The three cognitive
load levels-low,
high, and medium-
cannot be separable
linearly within the
data set.
performance for all models varies
based on the user, reflecting the
subjective nature of EEG data. For
instance, Users 1 and 6 perform
best for all of the classifiers. It
could be the hypothesis that those
particular users being comfortable
with the MATB-II environment. In
contrast, the other participants
were new to this setup, and that's
why they perform differently for
different models. However, the TL
method remains better than 98%
for all users, while the accuracy of
baseline methods varies dramatically. Also, note that the
variance of the TL method is minimal, and there is no
overlap in the notch of TL boxplot with any baseline methods,
which proves that the TL method is significantly better
than all of the baseline methods.
100
100
60
70
60
70
123456 7
User ID
Baseline_LDA
Baseline_RF
Baseline_LDA
Baseline_RF
Baseline_NN
Baseline_Logistic
Baseline_NN
Baseline_Logistic
TL
123456 7
User ID
Baseline_LDA
Baseline_RF
Baseline_LDA
Baseline_RF
Figure 6. The recall of the TL and baseline methods. The
figure is drawn using the methods described in the " TL
Method " section. The recall for all models varies based
on the user, which reflects the subjective nature of EEG
data. However, the TL method remains much better
than the baseline methods. Micro averaging is used for
precision, recall, and f1-score computation.
Baseline_NN
Baseline_Logistic
Baseline_NN
Baseline_Logistic
TL
Figure 7. The f1-score of the TL and baseline methods.
The figure is drawn using methods described in the
" TL Method " section. The f1-score for all models varies
based on the user, which reflects the subjective nature
of EEG data. However, the TL method remains much
better than the baseline methods. Micro averaging is
used for precision, recall, and f1-score computation.
Table 2. The results for TL and baseline methods for cognitive load identification.
Accuracy
Precision
Baseline
LDA
Logistic Regression
NN
RF
TL
0.7449 ± 0.05
0.7231 ± 0.09
0.8692 ± 0.06
0.9290 ± 0.03
0.9878 ± 0.01
0.7470 ± 0.05
0.7230 ± 0.09
0.8781 ± 0.05
0.9304 ± 0.03
0.9881 ± 0.01
Recall
0.7449 ± 0.05
0.7231 ± 0.09
0.8692 ± 0.06
0.9290 ± 0.03
0.9878 ± 0.01
F1-Score
0.7449 ± 0.05
0.7222 ± 0.09
0.8682 ± 0.06
0.9290 ± 0.03
0.9878 ± 0.01
January 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 13
Recall (%)
F1-Score (%)
IEEE Systems, Man and Cybernetics Magazine - January 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2022
