In this experiment, the potential of PPG signals was examined in recognition of 14 affective states. The PPG signals of DEAP database were chosen. The novel PPG descriptors, ECWS, was introduced. The recognition was done using PNN. Previously, in emotion recognition literature, PNN has been successfully applied on physiological signals (Goshvarpour et al., 2017a; 2017b). In addition, the influence of changing the sigma parameter in the performance of the classifier was evaluated.
As shown in Figure 4, the maximum accuracy rate of 100% was achieved using sigma 0.001 to 0.301. Discounting the sigma parameter, the classification rates were in the range of 74.37 to 100% for classifying 14 emotion classes.
Table 2 shows the comparative results of the proposed framework with the studies on automatic emotion recognition using PPG signals. Verhoef et al. (2009) used PPG and GSR morphological indices to design emotion recognition. For GSR, the number of responses means amplitude, mean rising time, and the mean energy of the responses was extracted. For PPG, heart rate variability (HRV), the mean interbeat interval (IBI), the standard deviation of the IBI, and the mean amplitude of the IBI were computed. Using a static Bayesian network, the highest performance of 60% was reported in the classification of 7 emotion categories. Koelstra et al. (2012) examined the multi-modal DEAP signals of 32 subjects to classify valence and arousal categories. They fed the spectral power asymmetry in the four EEG bands into the Fisher and Naïve Bayes. The valence and arousal based emotions were categorized with the rate of 57.6 and 62%, respectively. Park et al. (2013) used ST and PPG signals of 5 volunteers to classify happiness and sadness. A time interval between two successive PPG peaks and ST amplitude were appended by SVM. It has been reported that the accuracy rate of ST was 89.29%, the recognition rate of PPG was 63.67%, and a combined feature resulted in the highest recognition rate of 92.41%. An emotion recognition scheme was proposed by Li et al. (2014) using statistical features of ECG, GSR, and PPG. MLP was able to recognize 4 categories of emotion with the highest rate of 78.06%. Recently, Khan and Lewo (2016) used a ready-made platform for PPG and GSR. They tested two classifiers including decision tree (J48) and IBK to categorize 8 classes of emotion. The maximum accuracy was about 92%.