Channel-Reduced Hybrid Deep Neural Networks for High-Accuracy EEG-based Emotion Recognition

Abstract

Purpose: Emotion detection using electroencephalogram (EEG) signals has attracted increasing research attention due to its potential applications in affective computing and brain–computer interfaces. However, the inherently non-linear and non-stationary nature of EEG signals poses major challenges for designing a deep network framework capable of achieving high recognition accuracy. To address this, we propose a hybrid deep learning architecture that integrates one-dimensional convolutional neural networks (1D-CNNs) with gated recurrent units (GRUs) and long short-term memory networks (LSTMs) to enhance feature extraction and emotion classification.

Materials and Methods: Experiments were conducted on the DEAP dataset, with a special emphasis on decreasing the number of EEG channels from the full set of 32 to 8 and 4, while maintaining robust performance. CNN layers were utilized to capture spatial dependencies, whereas GRU and LSTM layers modeled temporal dynamics. To optimize performance, the Adamax optimizer and exponential linear unit (ELU) activation function were applied, ensuring stable convergence and efficient training.

Results: For an eight-channel EEG configuration, the proposed 1D-CNN-LSTM/GRU models achieved accuracies of 99.92% and 99.82%, respectively. When the input was further reduced to four channels, the models maintained a strong classification capability, achieving accuracies of 96.38% and 96.53%, respectively. These findings demonstrate that a considerable reduction in EEG channels does not necessarily compromise recognition performance.

Conclusion: The proposed hybrid framework successfully balances accuracy and efficiency, showing that EEG-based emotion recognition can achieve near-perfect performance with fewer channels. This reduction lowers computational cost, simplifies system design, and enhances practicality for real-world applications. Future work will involve applying the framework to alternative datasets and exploring adaptive channel selection strategies to further generalize its applicability.