Combined Weighted Feature Correlation Approach for Enhanced EEG-Based Emotion Recognition Across Diverse Datasets.

Area of Science:

• Affective computing and neuroscience
• Brain-computer interfaces
• Biomedical signal processing

Background:

• Emotion recognition from electroencephalogram (EEG) signals is crucial for understanding brain processes.
• Existing research often relies on linear EEG features, potentially missing finer emotional changes.
• Non-linear features offer a more nuanced approach to characterizing emotional responses.

Purpose of the Study:

• To investigate the impact of non-linear EEG features on emotion detection performance.
• To introduce and evaluate the Combined Weighted Feature Correlation (CWFC) model.
• To enhance emotion recognition accuracy using Generative Adversarial Network (GAN) data augmentation.

Main Methods:

• EEG data preprocessing with bandpass filters to isolate frequency bands (beta, alpha, gamma, delta, theta).
• Feature extraction using Independent Component Analysis (ICA), Discrete Wavelet Transform (DWT), and Fast Fourier Transform (FFT).
• Implementation of the CWFC model, Random Forest classifier, and GAN data augmentation, with an emphasis on Long Short-Term Memory (LSTM) for feature selection.

Main Results:

• The CWFC model with GAN data augmentation achieved 88% valence and 86% arousal accuracy on the DEAP dataset.
• Without GAN augmentation, the model reached 62% valence and 65% arousal accuracy on the DEAP dataset.
• The model demonstrated 89% average accuracy in distinguishing neutral, negative, and positive emotional states on the SEED dataset.

Conclusions:

• The proposed CWFC model, enhanced by GAN data augmentation, significantly improves EEG-based emotion recognition accuracy.
• Non-linear EEG features are superior to linear features for capturing intricate emotional patterns.
• The method shows robust performance across different datasets, highlighting its potential for real-world applications.