Attention-augmented CNN-LSTM framework for EEG-based spatiotemporal emotion recognition

Abstract

EEG emotion recognition offers a direct path to the analysis of affective brain states. This work presents an attention-enhanced CNN-LSTM model for emotion recognition from EEG in the SEED benchmark dataset (three classes: positive, neutral, negative). The suggested method combines Power Spectral Density (PSD), Differential Entropy (DE), and hemispheric asymmetry descriptors (ASM, DASM, RASM) into a single input description, which is trained through a hybrid architecture in which a CNN learns the patterns of spatial channels, stacked LSTM layers learn the patterns of temporal dependencies, and a temporal attention block highlights the most informative time segments. The model has an accuracy of 98.63% in 5-fold cross-validation across subjects, with averaged precision, recall, and F1-score exceeding 98.6. Compared with solutions from representative published SEEDs listed in Table 1, these findings indicate competitive performance and underscore the usefulness of complementary spectral and hemispheric signal fusion before attention-directed spatiotemporal modeling.