Emotional state is often assessed using electroencephalography (EEG), a non-invasive neuroimaging technique that measures brain electrical activity. Machine-learning approaches to EEG-based emotion recognition have been explored, but regression models for predicting emotion intensity remain unexplored. Moreover, few EEG-derived involvement indexes, defined as ratios of the spectral powers of two or more EEG rhythms, have been used for emotion recognition. This study aims to use EEG-derived involvement indexes to develop predictive models capable of identifying and rating specific emotional states, introducing a novel methodological framework that integrates 37 involvement indexes together with the powers of standard EEG rhythms as interpretable input features for predicting emotion intensity. To do so, 64-channel EEG signals recorded from healthy subjects while watching emotionally evocative videos were preprocessed using EEGLAB. Then, 6 EEG rhythms and 37 involvement indexes were extracted and given as inputs to five separate regression models, each corresponding to one target emotion (i.e., Anger, Sadness, Happiness, Disgust, and Fear), implemented using LASSO regression. Model performance was evaluated via mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R 2 ). Results showed MAE values between 0.90 and 1.19, RMSE between 1.23 and 1.56, and R 2 between 0.79 and 0.81, with Anger predicted most accurately and Happiness showing the highest variability. These findings demonstrate that combining the power of EEG rhythms with involvement indexes enables reliable and interpretable prediction of emotion intensity, highlighting the value of LASSO regression in handling high-dimensional EEG data.
Dell’Orletta et al. (Tue,) studied this question.