*3.4. Feasibility*

Most previous studies on emotion classification used multiple EEG channels. The feasibility of emotion classification using a single-channel in-ear EEG should be evaluated first. The feasibility evaluation was conducted by performing an emotion classification experiment using secondary data from the Dataset for Emotion classification using Physiological and Audiovisual Signals (DEAP) [47]. DEAP data set is a publicly available dataset for Brain Computer Interface (BCI) based emotion study provided by Koelstra S., et al. [47]. 32 channel EEG data from 32 subjects was collected, while they watched music video clips that were chosen to elicit emotions. The emotions elicited were based on the valence–arousal model. Valence was associated with emotion positivity which ranged from unpleasant to happy/pleasant. Arousal was associated with excitement which ranged from calm to excited. The subjects rated the music video clips on valence–arousal scales. The DEAP dataset was hence labelled, and the classification accuracy on the data could be evaluated by the subjects' rating. Out of 32 channels, only T7 and T8, which were stated to be close and correlate to the in-ear EEG were

used for our emotion classification. Our emotion classification using DEAP dataset will be used for evaluating and comparing to the in-ear EEG emotion classification accuracy.

Support vector machine (SVM) which was widely used for emotion classification [1,7,10,16] was used as a classifier. SVM has good generalization and overfitting prevention properties. Therefore, it is considered suitable for this work. Six statistical parameters by Picard et al. [48] were used for signal feature extraction ona3s time-lapsed window. The Butterworth filter was used to notch 50 Hz noise, and filter EEG signals into five frequency bands; namely, delta, theta, alpha, beta, and gamma bands [6]. Ten-folded cross validation was applied to suppress biases [49].

#### *3.5. Experiment Setup*

This experiment was designed to collect EEG data using our in-ear EEG electrodes when subjects' emotions were stimulated by pictures and sounds, described in Section 3.3. The results would be analyzed to assess the performance of in-ear EEG on emotion classification.

Twelve male and one female subjects aged between 20 to 30 years with an average age of 24, were recruited for emotion classification experiments. Before the experiment started, the impedances of the in-ear EEG were re-measured as quality assurance. An in-ear EEG device was then inserted into either the right ear or left ear according to each subject's preference, whereas earphones were inserted into the other ears. Earwax was cleaned by alcohol before the in-ear EEG insertion.

Unless the subjects preferred to put the in-ear EEG on the left, it was put on the right ear as the left ear is shown to be better for listening to music [50]. The ground electrode was placed at forehead and the reference electrode was placed at either cheek inferior to the ear. A small amount of saline was used as electrolyte gel. Forty trials were recorded per subject. IAPS and GAPED pictures were randomly displayed to the subjects. The total number of pictures used for each emotion was as suggested by IAPS and GAPED datasheets.

Each picture was displayed for 30 s. Subjects were recommended not to move during each picture viewing. Fifteen seconds of black screen was displayed after each picture in order to neutralize subjects' emotions before the next picture was displayed. During the black screen subjects were free to mobilize. After eight pictures, subjects could have a small break and were free to move around before they were ready to continue.

After the experiments were finished, the subjects were asked to evaluate their emotional response on each picture for emotion classification. This is because the emotional response to each picture may be di fferent among subjects or di fferent from the IAPS and GAPED datasets.

Statistical analyzes for any group comparison were performed using either *t*-tests or ANOVA, depending on the number of groups. A *p*-value of less than 0.05 was considered statistically significant. All statistics were performed using SPSS (IBM Corp., New York, USA)
