Prediction of Arousal and Valence State from Electrodermal Activity using Wavelet based ResNet50 Model

Abstract

Real time physiological signals like Electrodermal Activity (EDA), measured using wearable devices, could be used for emotion recognition, thus enabling new paradigms of health monitoring, including mental health. This paper analyses EDA signals, a measure of sympathetic nervous system activity, which plays a significant role in emotional regulation. A public database was used to categorize emotions based on valence and arousal values. The Continuous Wavelet Transform (CWT) was used to obtain the scalograms of the signals. The scalograms were then fed into the deep learning models, VGG16 and ResNet50, to recognize emotions. VGG16 gives a Pearson Linear Correlation Coefficient (PLCC) value of 0.7199 for valence and 0.7593 for arousal, with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values of 0.7653 and 1.0284, respectively. ResNet50 performs better, with a PLCC value of 0.7763 for valence and 0.8207 for arousal, and lower MAE and RMSE values of 0.7158 and 0.8712, respectively. This study proposes an adapted ResNet50 model for emotion recognition which could be integrated into wearable devices as a prospective feature.