Classification of Human Emotions Based on Multivariate Spectral EEG Attributes and CNNs

Human emotions are complex behavioral processes involving neural and chemical interactions in many layers of the human body. It plays a significant role in physical and mental health, in daily decisions, and in social communication. Investigating emotion is among the most popular areas of electroencephalogram (EEG) topics in the literature since physiological signals are hard to conceal and it provides harmless recording, easy accessibility, portability, and cost-effectiveness. Hereby, we present a transfer learning method to classify human emotions by EEG. First, raw EEG recordings were preprocessed and segmented. Then, the multivariate synchrosqueezing transform (MSST) was applied to these segments, and the obtained time-frequency images were fed into a CNN-based transfer learning model. In total, 45 EEG recordings from 15 subjects were trained across repetitive sessions. Inception V3 architecture was used in the training. Performance analyses were conducted for a binary classification that separates positive and negative emotions. In the leave-one-subject-out training approach, up to 100% validation accuracy was reached across 3 different window sizes. Among all window size-based training, it was observed that the 16-sec window size showed 78.05% average validation accuracy. The results proved that the proposed method demonstrated highly promising results in subject-based emotion classification. However, the subject-based average accuracy was not as high as the training of the shuffled dataset.