Advancing EEG Research on Human Emotions Through Deep Learning Models

Abstract

Emotions are critical in driving human cognitive and social activities. As a non-invasive technique for recording neuronal electrical activities, electroencephalography (EEG) has been widely used in diagnosing and treating conditions such as epilepsy and depression since Hans Berger first recorded human brainwaves in 1924. It has also gradually expanded into human-computer interaction applications. However, emotion research requires large-scale annotated data, while challenges such as low signal-to-noise ratio and significant individual variability in EEG hinder real-time, high-precision emotion recognition. Ethical and privacy concerns further restrict its application. This paper systematically reviews recent advances in deep learning models for EEG-based emotion analysis to address these issues. It proposes an end-to-end processing framework centered on deep neural networks. The study notes that classical deep learning algorithms, relying on the independent and identically distributed assumption, struggle to adapt to emotional variations across individuals and scenarios. In contrast, models like convolutional neural networks and generative adversarial networks (GANs) can automatically perform noise reduction, artifact removal, and frequency-band feature extraction during preprocessing, significantly reducing manual intervention. Furthermore, a "two-stage classification" strategy is proposed: first extracting coarse-grained emotion labels (e.g., happiness, anger, sadness) in time and frequency domains, then further refining emotions based on emotion-specific frequency bands (e.g., "mild pleasure" vs. "extreme excitement"). Dynamic emotional changes are captured using methods such as the short-time Fourier transform and the continuous wavelet transform. Finally, the paper discusses unresolved challenges, including data scarcity, real-time processing, privacy security, and ethical compliance, suggesting future adoption of federated learning for enhanced implementation.