Classification of Emotions (Positive-Negative) Based on EEG Statistical Features using RNN, LSTM, and Bi-LSTM Algorithms

Affective computing research related to EEG-based emotion recognition has become a current research trend. This research becomes very interesting because the EEG signal is complex and always changes depending on the condition of the individual at that time. So, if the information in the EEG signal can be extracted, a person’s emotional state (which tends to be hidden) will be revealed. Therefore, this study directly proposes an automatic emotion recognition system with recorded EEG data. In this study, EEG recording was performed on 32 participants. Raw EEG data is processed by stages such as pre-processing, subband decomposition, feature extraction, and classification of emotions based on feature values. The EEG signal features explored include mean value, MAV, standard deviation, variance, skewness, kurtosis, zerocrossing rate, and median. Based on the results of EEG feature extraction, it can be seen that positive-negative emotions have different feature values and the differences are also significant. The results of signal feature extraction are presented based on channels (FP1, FP2, F7, and F8) and EEG subbands (Alpha, Beta, and Gamma) for each emotional state (positive-negative). In addition, the best accuracy values for emotion classification are 93.75% (RNN), 93.75% (LSTM), and 92.97% (Bi-LSTM) in the classifier testing process.