Assessment of psycho-emotional state of a person using AI-based EEG analysis
Keywords:
Abstract
Introduction: Determining emotional valence based on electroencephalographic (EEG) data is an urgent issue. Nonetheless, traditional multi-electrode EEG systems are unsuitable for everyday use. In addition, signals from portable, low-channel devices exhibit high noise levels and significant variability across users, complicating their analysis and interpretation. Purpose: To develop a deep-learning-based approach for EEG data analysis suitable for a reliable assessment of emotional valence using low-channel wearable devices. Results: We propose a comprehensive approach combining signal processing with classification via convolutional and recurrent neural networks. Neural network models were trained on a publicly available multi-channel dataset, with the subsequent transfer to our recordings obtained with the use of a wearable four-electrode headband BrainBit. The experimental results demonstrate an emotion valence recognition accuracy (positive, negative, and neutral emotions) of 70–75% during cross-subject validation on public data, reaching 85–91% accuracy as compared to computer vision methods using our own data. Practical relevance: The findings confirm the feasibility of applying these models in wearable systems for monitoring human emotional and cognitive states. Discussion: There is a recognized need for further model adaptation to reduce inter-subject variability and noise, as well as for the refinement of signal synchronization methods and data labeling approaches.
