Statistical processing of facial electromyography (EMG) signals in emotional film scenes

To improve human-computer interaction, computers need to recognize and respond properly to their users’ emotional state. As a first step to such systems, we investigated how emotional experiences are expressed in various statistical parameters of facial EMG signals. 22 Subjects were presented with 8 emotional film fragments while a TMS Portilab system was used to measure the activity of frontalis above the left eye (EMG1), right corrugator supercilii (EMG2), and left zygomaticus major (EMG3). Additionally, subjects rated the intensity of both their positive and negative feelings for each film. Based on average positive and negative ratings, films were classified into 4 emotion categories (with 2 films each): Mixed, Neutral, Positive, and Negative. From each EMG signal, 6 statistical parameters were derived: mean, absolute deviation, standard deviation, variance, skewness and kurtosis. For each of the resulting 18 parameters, a REMANOVA was conducted, with the 4 emotions and 2 films as within-subject factors. The effect of emotion was significant for EMG2 skewness (F(3,18)53.500, p=.037), EMG3 mean (F(3,18)59.711, p<.001), EMG3 absolute deviation (F(3,18)58.369, p<.001), EMG3 standard deviation (F(3,18)55.837, p=.006) and EMG3 variance (F(3,18)54.064, p=.023). Thus, only few of the EMG parameters reached significance, possibly because mimicking a potential human-computer situation we did not correct our data for baseline values and averaged over a period as long as 120s. Nevertheless, the EMG3 signal remains promising in its differentiation among emotion categories