Discrimination between different emotional states based on the chaotic behavior of galvanic skin responses

Abstract

The purpose of the current study was to examine the effectiveness of galvanic skin responses (GSRs) in emotion recognition using nonlinear approaches. GSR of 35 healthy students was recorded while subjects were listening to emotional music clips. The signals were comprehensively characterized by nonlinear features. Applying three dimensionality reduction methods, including sequential forward selection (SFS), sequential floating forward selection, and random subset feature selection (RSFS) in combination with four classification approaches, including K-nearest neighbor, least-square support vector machine, Fisher discriminant analysis, and quadratic analysis, discrimination between emotional classes was evaluated. In addition, two classification strategies were examined, including binary (BIC) and one vs. rest. The results showed that higher recognition rates were achieved for Fisher. In this case, the BIC accuracy rates were higher than 99% in all emotional states and all feature selection methodologies. The maximum classification rate of 99.98% was obtained using RSFS and Fisher in sadness. Among all emotion categories, better recognition rates were achieved for peacefulness and fear. This study demonstrates that nonlinear GSR characteristics can provide an informative measure to investigate the physiological fluctuations in different emotional states during music.