Analysis of Dynamics of EEG Signals in Emotional Valence Using Super-Resolution Superlet Transform

Abstract

Electroencephalography (EEG)-based emotional state assessment is widely preferred due to its noninvasiveness and nonradiation approach. However, these signals are highly nonstationary and multicomponent, demonstrating large intrasubject variability. Extracting time and frequency information simultaneously from EEG addresses these challenges to effectively recognise the valence emotional states. Traditional time–frequency (TF) approaches optimise either temporal or frequency resolution, resulting in failure to identify fast transient oscillatory emotional events. In this letter, an attempt has been made to recognize emotional valence using super-resolution-based superlet transform (SLT). For this, the preprocessed EEG signals during emotion-evoking audio–visual stimuli from publicly available database is considered. The EEG signals are decomposed into theta, alpha, beta, and gamma frequency bands and are subjected to SLT. The TF skewness and kurtosis are extracted from the SLT. The statistical significance of features is evaluated, and the features are applied to three machine learning algorithms: random forest, Adaboost, and k-nearest neighbor. The results show that the SLT-based TF spectrum is able to provide variations of frequency components associated with emotional valence. Both the features exhibit statistically significant $(p < 0.05)$ difference in the high-frequency gamma bands to characterize emotional valence. Among the classifiers, AdaBoost stands out as the most robust performer (F1 = 70.16%). Feature importance analysis highlights that SLT features from the fronto-central and parieto-occipital brain regions play a crucial role in valence detection. It appears that this method could be useful in analyzing various mental well-being conditions in clinical settings.