EEG Emotion Recognition using Parallel Hybrid Convolutional-Recurrent Neural Networks

Electroencephalogram (EEG) signals of certain emotions contain waves with specific frequency bands. So, emotion recognition uses the network containing each wave to become relevant. EEG signals record electrical activity in the brain from several channels. Therefore, EEG signal processing needs consideration to spatial and temporal. Spatial is a signal between channels, while temporal is a sequence. Several methods were used, Convolutional Neural Networks (CNN) with various dimensions, Recurrent Neural Networks (RNN), and hybrid CNN-RNN. This paper proposed a hybrid 2D CNN-RNN method for identifying emotions from a parallel network of each wave. Two-dimensional CNN is used in channel extraction in a short time of the signal. Using short-time signals is intended to minimize the non-stationary characteristic of EEG signals. Meanwhile, the identification of emotions is carried out with RNN using the output of 2D CNN extraction. The modeling and testing used a dataset from SEED, with three emotion classes: positive, neutral, and negative. The experimental results show that using a split network of each wave increased accuracy from 80.92% to 84.71% and a decreased Loss value. While the use of 2D CNN only increased a less significant accuracy than 1D CNN. Evaluation of the waves shows that Beta and Gamma waves provided the best precision, 87-91%, and Theta waves gave 79-85% precision. Alpha wave degrades overall performance, which only has 56-61% precision, considering it is a mid-wave between Theta and Beta. It is necessary to choose the proper weight updating technique. Adaptive Moment (Adam) increased accuracy than AdaDelta, AdaGrad, and RMSprop.