A Deep Learning Approach for Classifying Emotions from Physiological Data

Affect detection form physiological signals has been an active research field recently. It has many applications in education, computer games, car driver monitoring, and health. Electrocardiography, electromyography, respiration, and skin conductivity have all been considered as potential physiological channels for recognizing affective states. There is a large amount of literature on physiology-based affect detection that reports on high detection rates of affective states (either discrete emotions, or dimensional valence/arousal) under laboratory conditions. However, one challenge is that physiological measures exhibit considerable variations over time, non-stationary, making classification of future data difficult. In this study, we aim to investigate the changing physiological features over multiple sessions. In our experiments, we use a dataset collected from four participants who viewed a set of images from the International Affective Picture System (IAPS) over multiple days. This study also provides a comparison between self-reported and IAPS ratings for both valence and arousal dimensional affective model. In addition, the study provides an analysis of the performance of individual physiological channels for affect detection. Results indicate that a Deep Neural Network achieved accuracy with 61.12%, 67.14% of detecting self-report compared with IAPS rating for valence and arousal. Results also show that EMG zygomatic features and a fusion of features from all physiological channels have the highest affect detection accuracy for both valence and arousal.