Convolutional-LSTM Network for Emotion Recognition Using EEG Data in Valence-Arousal Dimension

The detection of emotions using automatic electroencephalogram (EEG) analysis is a significant challenge within human–computer interaction. This study aims to present a hybrid model that considers the frequency and time characteristics of multimodal EEG data to facilitate emotion recognition. Most conventional methods for identifying emotions rely on analyzing the frequency characteristics of EEG data. Spatial qualities are advantageous as they encompass data about a person’s affective state. This study analyzes EEG signals and utilizes deep neural networks for feature learning and end-to-end categorization. To begin with, the initial step involves converting one-dimensional raw EEG signals into a scalogram using continuous wavelet transform (CWT). The scalogram is an input for the Convolutional-LSTM framework, which extracts features and classifies affective states. This framework operates on multiple-channel EEG data, considering temporal and spectral information. To evaluate the effectiveness of our proposed model, we conducted experiments using the DEAP dataset, a well-established benchmark dataset for emotion recognition. The experimental results demonstrated that the suggested model attained classification accuracies of 57.47 and 60.17% for valence and arousal, respectively. The proposed methodology suggests that the scalogram technique efficiently determines an individual’s emotional states based on EEG signals.