A novel variational nonlinear chirp mode decomposition-based critical brain-region investigation for automatic emotion recognition

Abstract

The field of affective computing that deals with emotion recognition from physiological information, particularly electroencephalography (EEG), is becoming more and more important. Researchers have employed these signals to identify emotions, but the majority of them only distinguish a limited number of emotional states (e.g. happiness, fear, sadness, calmness, anger, etc.). This study outlines the proposed model for classifying nine emotional states from the DREAMER database into arousal, valence, and dominance dimensions using a discrete scale. To create it, we studied variational non-linear chirp mode decomposition (VNCMD) to decompose the EEG signals into five distinct modes. 11 entropy features were extracted from each mode. Then a comparative analysis of six popular machine learning-based classifiers was performed. The VNCMD's third mode with a rotational random forest classifier gave the best performance parameters for this application. This systematic analysis revealed that the best brain region, i.e., the anterior frontal area using only 2 channels, provided the highest accuracy and F1-score for arousal: 96.07% and 97.30%, valence: 94.49% and 95.58%, and dominance: 95.83% and 97.19% categories. Results, using the DREAMER dataset, show that our model can have the highest predictive ability with AUCs of 0.98, 0.98, and 0.97 for the arousal, valence, and dominance categories, respectively. The findings of this study demonstrate that the features, brain regions, and machine learning models commonly employed in emotion classification tasks may be used in more difficult tasks, like the prediction of precise values for arousal, valence, and dominance.