Machine learning algorithms for emotion recognition through the analysis of eye-tracking data in a sample of healthy subjects

Emotion recognition is a research field of major interest to areas such as health care and psychology. Many emotion models and emotion recognition algorithms have been developed in the last decades to identify and better understand human emotions. Their recognition can be attempted through the analysis of physiological responses to specific emotional stimuli; among others, oculomotor activity and pupil dilation responses have been repeatedly investigated in the literature, due to its low cost and non-invasive nature. This work investigates, in a sample of 15 healthy participants, differences in these responses to customized emotional audio-visual stimuli delivered through a virtual reality headset, consisting of pre-recorded videos featuring either acquaintances or strangers highlighting positive characteristics and future plans of either the persons under study or other individuals, respectively. To this aim, eye-tracking data has been acquired, pre-processed and a feature extraction procedure has been achieved. Subsequently, machine learning algorithms have been used, first, to differentiate between stimulation and non-stimulation periods; second, to differentiate between stimuli featuring acquaintances and strangers; third, to differentiate between stimuli of specific self-reported pleasure (i.e. valence) and excitation (i.e. arousal) values. Our findings showed an accuracy higher than 90% at differentiating stimulation and non-stimulation periods, and an accuracy around 80% at differentiating videos from acquaintances and strangers. The accuracy at identifying stimuli with specific characteristics was higher than 65% and comparable to the accuracy reported by previous studies. In light of these results, additional studies are needed including more participants and a wider range of emotions to get more generalised and accurate results.