Emotion Classification Through EEG Signals: The EmoSense Model

In the Emotion research, there has been extensive work on emotion classification using electroencephalogram (EEG) signals. Despite most studies having achieved high accuracy, there is a potential for improvement in the model, particularly in training time reduction to optimize the modeling efficiency. Hence, our study aims to develop an efficient EEG-based emotion classification model that achieves high accuracy with a reduced number of training epochs. The proposed model, named EmoSense, employs a hybrid deep learning neural network architecture, which comprises a combination of Convolutional Neural Network (CNN), Deep Autoencoder (DAE), Bidirectional Long Short-Term Memory (BiLSTM), and an Attention mechanism. EEG features are first extracted using Differential Entropy (DE) on the EEG channels, which are then structured to form a four-dimensional (4D) representation of features with temporal, spatial, and frequency information. Subsequently, the extracted features will be used as input to the proposed model for training. Experimental evaluation of the publicly available DEAP dataset demonstrated that the proposed model achieved an accuracy of 94.11% with only 35 epochs. This performance is comparable with the result of the Cascaded Convolutional Recurrent Neural Networks model developed by Meng et al., which reported an accuracy of 94.64% with 100 epochs. The primary reason for the comparison with Meng et al. is that our overall process, including feature extraction and formation, is similar to their study, with the only difference being the deep learning architecture. Our finding indicates that the EmoSense model is not only able to achieve comparable accuracy but also reduces the number of training epochs, demonstrating the potential to learn the EEG features in a shorter time and lowering the computational cost. This study will benefit EEG-based emotion recognition to be more accessible and practical for real-world applications.