Comprehensive Analysis of Feature Extraction Methods for Emotion Recognition from Multichannel EEG Recordings

Area of Science:

• Neuroscience
• Computer Science
• Signal Processing

Background:

• Electroencephalogram (EEG)-based emotion recognition is advancing due to signal processing and machine learning.
• Previous studies often used limited data and EEG features, hindering direct comparison and validation.

Purpose of the Study:

• To comprehensively compare the classification accuracy of diverse EEG feature sets for emotion recognition (valence and arousal).
• To evaluate feature set performance across multiple independent datasets for robust validation.

Main Methods:

• Investigated five EEG feature sets: statistical, fractal dimension (FD), Hjorth parameters, higher-order spectra (HOS), and wavelet-derived features.
• Evaluated performance using Support Vector Machine (SVM) and Classification and Regression Tree (CART) classifiers.
• Utilized five public EEG datasets (MAHNOB-HCI, DEAP, SEED, AMIGOS, DREAMER) for cross-dataset validation.

Main Results:

• The Fractal Dimension (FD) with CART classifier achieved the highest mean classification accuracy: 85.06% for valence and 84.55% for arousal.
• Consistent performance across all five datasets indicates the reliability of FD features for emotion recognition.
• This study provides a comparative analysis of various EEG features for emotion detection.

Conclusions:

• Fractal dimension (FD) features are highly effective and reliable for EEG-based emotion recognition.
• The findings pave the way for developing real-time EEG-based emotion recognition systems.
• This research offers a robust framework for feature selection in affective computing.