Work

Optimizing Gated Recurrent Unit Architecture for Enhanced EEG-Based Emotion Classification

Year: 2025
Type: article
Abstract: Emotion recognition using EEG signals has gained significant attention in affective computing and brain-computer interface (BCI) applications. However, achieving high classification accuracy remains a challenge due to the complexity and variability of EEG signals. This study aims to optimize the Gated Recurrent Unit (GRU) model for improving the performance of EEG-based emotion classification. The approach involves feature selection and architectural modifications to the GRU model. Selected EEG features include mean, standard deviation, statistical moments (skewness and kurtosis), min-max values, logarithmic covariance matrix, covariance matrix, Shannon entropy, log-energy entropy, Fast Fourier Transform, and autocorrelation, extracted from alpha and beta frequency bands. The proposed GRU model consists of four stacked GRU layers with decreasing hidden state sizes, ensuring efficient temporal feature extraction while reducing computational complexity. Experimental results demonstrate the superiority of the proposed GRU model compared to Simple RNN, LSTM, and traditional Machine Learning models (Naïve Bayes, SVM, Random Forest, and Linear Regression). The GRU model achieves high recall (98.81%), specificity (99.42%), precision (98.82%), accuracy (99.22%), and F1 Score (98.81%), outperforming alternative models in all evaluation metrics. These findings indicate that the GRU model effectively captures temporal dependencies in EEG signals, making it a robust and efficient approach for EEG-based emotion classification. In conclusion, this research confirms that GRU is an optimal deep learning model for emotion recognition using EEG. Future research could explore multi-modal emotion recognition, attention-based architectures, and real-time deployment in wearable EEG devices to further enhance classification accuracy and real-world applicability.
Language: English
Open Access status: diamond
OpenAlex ID: W7160911007
FWCI: 0.245
Cites: 10
Cited by: 1
Related to: 10