EEG emotion recognition method integrating improved dynamic graph convolutional neural networks and bidirectional simple recurrent units

Abstract Electroencephalogram (EEG), as a core physiological indicator for emotion recognition, suffers from problems such as poor adaptability of static modeling and inefficient temporal capture due to its non-stationarity, high dimensionality, and multi-domain dispersion of emotional information. To propose a high-precision and high-efficiency EEG emotion recognition method and overcome the current technical bottlenecks, this paper studies and constructs a three-level architecture of "multi-domain feature extraction—dynamic spatial modeling—efficient temporal learning". First, multi-domain brain map features are extracted through a time–frequency spatial domain fusion strategy. Key features are then enhanced by an attention module. Finally, a hybrid model of dynamic graph convolutional neural network (DGCN) and bidirectional simple recurrent unit (BiSRU) is constructed to fuse channel relationships (CR). Validation results on the DEAP and SEED datasets show that the hybrid model achieves an accuracy of 95.2% and an area under the receiver operating characteristic (AUC) of 0.978 on the DEAP dataset, and 98.1% and an AUC of 0.992 on the SEED dataset. In practical applications, the model achieves an accuracy of 93.7% in classifying emotions among university students and 89.8% in predicting abnormal emotions among corporate employees. These findings reveal that the model significantly improves recognition accuracy and real-time performance through dynamic topology adaptation and efficient bidirectional temporal learning. This study provides a physiologically appropriate solution for scenarios such as mental health monitoring and intelligent healthcare, promoting the practical application of EEG emotion recognition technology.