Cross-Subject EEG-Based Emotion Recognition Through Neural Networks With Stratified Normalization

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Emotion recognition from physiological signals is a growing field with numerous applications.
• Electroencephalography (EEG) is a promising non-invasive and low-cost method for emotion detection.
• Inter-participant variability in EEG data necessitates complex calibration for cross-subject emotion classification.

Purpose of the Study:

• To introduce a new participant-based feature normalization method called stratified normalization.
• To address the challenge of inter-participant variability in EEG-based emotion recognition.
• To improve the performance of deep neural networks for cross-subject emotion classification.

Main Methods:

• Developed and applied stratified normalization for feature normalization in deep neural networks.
• Utilized the SEED dataset, comprising 62-channel EEG recordings from 15 participants.
• Compared stratified normalization with standard batch normalization for training.

Main Results:

• Stratified normalization significantly outperformed standard batch normalization in cross-subject emotion classification.
• The multitaper method for EEG feature extraction yielded the highest performance.
• Achieved 91.6% accuracy for binary (positive/negative) and 79.6% for ternary (positive/negative/neutral) emotion classification.

Conclusions:

• Stratified normalization effectively reduces inter-participant variability while preserving emotion-related information in EEG signals.
• The proposed method offers significant benefits for developing robust cross-subject EEG-based emotion recognition systems.
• This research highlights the potential of stratified normalization for advancing affective computing.