Table 2 shows an example of extracted feature values (scalp topography) of the dataset. A total of twelve such tables have been generated by extracting the corresponding twelve features. In Table 2, each row corresponds to an EEG epoch, whereas each column represents the feature-value of an EEG channel, and the last column labels each epoch as a blink (0) or a non-blink (1) category. Each of the five classifiers has been trained and tested with all the twelve features consecutively.
Comparison Of Features (Irrespective Of Classifiers)
Figure 5(a)-5(d) illustrate the comparative performances of the twelve EEG features based on accuracy, precision, recall, and f1-score as performance measuring metrics. The part of the figure for a particular performance measuring metric shows its’ highest scores, irrespective of classifiers, for all the twelve features.
For instance, the accuracy scores of scalp topography are (0.94, 0.88, 0.94, 0.92, and 0.97) % obtained with the five classifiers SVM, LR, KNN, NB, and ANN, respectively. Therefore, the best score, i.e., 0.97%, is shown in Fig. 5(a) as the accuracy score of scalp topography. Figure 5a shows the top accuracy scores of all the features. Here, scalp topography achieved the best score (0.97%) among all features. Max and peak-to-peak amplitude individually gained the second highest (0.96%) score close to the best score. The lowest score is 0.68% which is acquired by the feature mean.
According to the precision scores of Fig. 5(b), PSD, scalp topography, and max individually outperform by achieving 0.98%. The second highest score is 0.97% which is achieved by variance. Mean has been found as the worst feature scoring 0.73%. Several features individually achieved the top recall value (0.99%). These features are- scalp topography, variance, standard deviation, peak-to-peak amplitude, entropy, and max. The lowest recall value is 0.80%, which is gained by skewness.
While considering the F1 score, PSD, max, scalp topography, and peak-to-peak amplitude are the best features scoring 0.97%. The second-best feature is the standard deviation (0.96%). The worst feature is the mean by gaining 0.68%.
Comparison Of Classifiers (Irrespective Of Features)
For classifier comparison, in how many cases (features) a classifier achieved the highest score regarding a particular performance measuring metric is considered. In Fig. 6(a-d), the Y-axis value indicates the number of features for which a classifier scored best.
Figure 6a shows the comparison of classifiers based on accuracy. Here, ANN consistently scored top values for all the twelve features. SVM got the top accuracy score only with PSD, while other classifiers have no top accuracy score across any feature.
Figure 6b shows the classifier comparison under precision score. In this case, ANN got the top precision score across 07 features (PSD, max, mean, entropy, peak-to-peak amplitude, standard deviation, variance).
SVM achieved for six features, and LR, and NB achieved for 2 and 4 features, respectively. KNN has no top precision score with any feature.
While considering recall (Fig. 6c), KNN outperforms by getting the top score with 09 features. SVM and NB got top precision with 02 features, but LR and ANN have no success in this regard.
KNN again outperforms considering the F1 score. It achieved a top F1 score with 07 features. However, ANN, SVM, and NB achieved top for 5, 2, and 1 feature, respectively. LR has no achievement here.
Overall Result
To find out the best feature-classifier combination, the summation of the four performance measuring metrics scores of each feature-classifier combination is considered. The summation scale ranges from 0 to 4. Accordingly, scalp topography-ANN have found as the best combination, and their summation score is 3.87.
Figure 7 summarizes the overall results in a graphical view. The X-axis represents the features, whereas each column represents a classifier, and the stacked bars show the feature-classifier performances of the corresponding performance measuring metrics. Each Feature on X-axis takes five consecutive columns for the five classifiers- SVM, LR, KNN, NB, and ANN, respectively.