Human Emotion Detection Using Eeg Signals: Insights Into Feature Selection

Human emotions are complex phenomena and are closely linked with the physical and mental health of any individual. Therefore, accurate detection of human emotions is strongly relevant to quality of life and well-being. Brain signals such as electroencephalograms (EEG) in this regard are among the leading imaging modalities that can classify human emotions. However, EEG signals are complex, and EEGbased emotion recognition systems process hundreds of features, which makes EEG interpretation difficult and computationally complex. This is because not all features are equally important. Therefore, this study examines a pool of EEG features and investigates which features are essential for emotion detection. We developed a hybrid feature selection method that combines statistical, information-based, and tree-based approaches with brain research knowledge. Using the SEED dataset (310 features, 20 subjects, 41,160 samples), the method identified 29 important features. This is a 90.6 % reduction while keeping 88−92% of the original performance. Results show theta band waves (48 %) are the most useful features. Temporal brain regions (35 %) and frontal regions (28 %) are also important. Random Forest classification achieves 74.9% accuracy with the reduced features. It shows good performance across all emotion types (F1-scores: 0.805 negative, 0.613 neutral, 0.817 positive). The findings show that the hybrid approach performs well in balancing efficiency and accuracy and can potentially be extended to develop portable, real-time emotion recognition device in real-life conditions.