5.4 Results and Discussion
6.4.1 Event-Related Potential Classification
The single trial ERP detection method proposed by [56], is utilised to evaluate per- formance of the proposed system to remove eye-related artefacts. [13] showed in the captured dataset, strongest ERP responses are observed over occipital electrode sites. In order to avoid the curse of dimensionality, we used only 7 EEG channels placed over occip- ital sites for ERP analysis. First, eye-related artefacts are removed using our method and the method proposed by [13]; and cleaned EEG is reconstructed. Then, the cleaned EEG
Figure 6.5: Sample processed VOG data where missing values in (a) are substituted using interpolation; (b) shows the interpolated VOG data and the blink trigger signal derived from the missing values.
data is band-pass filtered (0.1−4Hz). The time samples between 50ms to 400ms after the stimulus, and 50ms to 400ms before stimulus onset are considered as, ERP and no-ERP features, respectively (see Figure 6.4). To remove baseline drift, prior to feature extrac- tion, the mean of the pre-stimulus segment (no-ERP) is subtracted from both the ERP and no-ERP segments. In order to reduce the dimensionality of data, after selecting ERP and noERP segments, the data samples are decimated with factor of 5 (for more details see [56]). Feature vector is made by concatenating ERP and no-ERP samples, extracted from each of the preselected EEG channels. In total 630 features (90samples × 7channels) are extracted. The extracted features are used to train a kNN classifier with Euclidean distance (the value of k is set to 1). The accuracy of ERP classification, for each subject, is obtained using a 10-fold cross validation. Therefore, all of the obtained accuracies are averaged over all subjects, and total accuracy of the system is obtained.
6.5
Results
Figure 6.5.b illustrates an instance where interpolation described in section 6.2.1 is per- formed to fill the VOG gaps in Figure 6.5.a. Additionally, Figure 6.5.b depicts the blink signal b(t), derived from 100 to 400ms gaps in the VOG data. As is shown in the fig- ure, the employed gap filling method avoided the occurrence of manipulated saccades by providing a smooth transition between the last data sample before the start of a gap and the first data sample after the end of that gap. Fig. 6.6 shows the distribution of the
Figure 6.6: Distribution of the Z-Score values obtained by cross-correlation of each single IC with b(t), x(t) and y(t). Scores belong to all subjects in (a) the Fixed SSVEP and (b) the Smooth Pursuit SSVEP experiments. Red lines indicate the selected threshold (Zγb,i
,Zγx,i or Zγy,i = 2.0)..
blink, horizontal and vertical eye movement Z-Scores of ICs of all subjects in fixed SSVEP (Fig. 6.6.a) and smooth pursuit SSVEP (Fig. 6.6.b). The ICs with Z-Scores higher than 2.0 are considered as EOG artefacts. As is shown in the figure, in the smooth pursuit SSVEP where there are intentional eye movements, the ICs considered as EOG artefacts are better separated from the mean distribution of all other ICs in the fixed SSVEP where subjects are instructed to avoid eye movements. It can suggest that ICA performed a bet- ter EOG source separation, when there are more eye movement activities conveying more information about behaviour of the EOG sources.
density of an IC automatically detected as the eye-related artefactual source. Considering the dipolar behaviour of the eyes, the distribution of the detected IC over the right and left frontal channels on the scalp map implies a horizontal eye movement. This is confirmed by rapid amplitude fluctuations, similar to saccade activity, in the IC’s time series; and higher spectral power of the IC below 5Hz, than all other frequencies.
The detected ICs are validated by comparing them to the ground truth. In the ground truth, from 140 ICs (2 experiments ×5 subjects ×14 ICs ) in the SSVEP dataset, 18 ICs are labeled as eye-related artefacts (8 blink ICs and 10 eye movement ICs).
Considering both fixed and smooth pursuit SSVEP, using the proposed method, there are 11 ICs detected as blinks, 10 ICs detected as horizontal eye movements and 10 ICs detected as vertical eye movement artefacts. The 10 ICs detected as horizontal eye move- ments are the same ICs detected as vertical eye movements, suggesting information of both types of eye movements are concentrated in the same ICs. Additionally, there are 2 ICs mutually detected as blink and eye movement artefacts. In total, there are 19 unique ICs detected as EOG artefacts (either blink or eye movement artefacts).
Table 6.1a shows the confusion matrix of the eye-related IC detection, using our method. Based on the confusion matrix, all of the 18 artefactual ICs labeled in the ground truth, are correctly detected using the proposed method. Additionally, the method labeled another IC which is not labeled as an artefact in the ground truth. The proposed method achieved BA = 97%, which indicates the reliability of the eye-related IC detection, based on the ground truth.
For comparison, Plo¨ch’s method is also applied to detect IC’s corresponding to eye- movements. Because Plo¨ch is not designed to detect blink sources, the ICs labeled as blink are excluded from the ground truth. Table 6.1b shows the confusion matrix of the eye movements source detection, using Plo¨ch method. Plo¨ch achieved BA = 69% for eye movement IC detection, 28% lower than our proposed method (p-value< 0.05).
(a) Saccade time Series
(b) Saccade PSD (c) Saccade topography
Figure 6.7: Typical example of (a) time series, (b) power spectral density and (c) scalp map (topography) of an IC, corresponding to the horizontal saccade, manually labeled in the validation procedure.