Deep Learning Analysis of Electrophysiological Series for Continuous Emotional State Detection

The detection of emotions has a wide range of applications in psychology, marketing, and human-computer interaction. The detection of emotional states using biomedical signals is attractive because it is non-invasive and there is a large body of knowledge in digital signal processing that can be applied to the analysis of these signals. Additionally, artificial intelligence can be used to improve the accuracy of emotion detection by using machine learning algorithms to analyze large amounts of data. In this study, we explored the application of a parallel hybrid architecture called the parallel TCN-SBU-LSTM, which combines a temporal convolutional network and a stacked bi-and uni-directional LSTM. For the EPiC 2023 Challenge, we set out to estimate the arousal and valence states of different subjects in four different scenarios, using a continuous analysis of five physiological signals. To determine the best hyperparameters for each scenario and which signals to use for estimation, we used a methodology involving grid search and 5-fold cross validation. The models obtained an average root mean square error of 1.585 across scenarios, demonstrating the suitability of the parallel TCN-SBU-LSTM network to estimate the emotions of the subjects in different scenarios with consistent performance.