A Multi-Scale Inverted Spatial-Temporal Network for EEG-Based Emotion Recognition

Abstract

Understanding human emotional states through Electroencephalography (EEG) signals has gained significant attention due to its applications in healthcare, human-computer interaction, and affective computing. However, existing approaches often struggle to model temporal dynamics and spatial dependencies effectively, which limits recognition accuracy. The primary research gap lies in the inability of conventional and recent models to simultaneously capture multi-scale temporal patterns while preserving channel-specific information over time. To address this limitation, this study proposes a MultiScale Inverted Spatial-Temporal Network (MIST-E) for EEG-based emotion recognition. MIST-E constructs multi-scale representations and employs an inverted embedding strategy to maintain temporal continuity and spatial channel relationships. In addition, a newly designed CNN is used to extract discriminative features for reliable classification. Experimental results on the DEAP dataset demonstrate that MIST-E effectively captures complex spatial-temporal dependencies, achieving 90.56±1.02% accuracy for valence and 91.12±0.98% for arousal. These findings indicate that MIST-E provides improved accuracy compared to existing methods.