MSCABiNET: A Lightweight Attention-Enhanced CNN-BiLSTM Model for EEG-Based Emotion Recognition

Neural signal-guided emotion analysis improves the use of algorithmic intelligence in the domains of behavioral computation and biomedical science, enhances comprehension of a person’s emotional states, and enables a better humancomputer interface. Conventional deep learning models utilized for emotion detection often suffer from high computational requirements and limited feature extraction abilities due to existing model architecture. The proposed MobileNet-SE-CoordAttention-BiLSTM (MSCABiNET) model addresses the limitation of existing emotion detection structure by introducing a lightweight, effective architecture that enhances spatialtemporal feature representation while reducing computational complexity. It develops a lightweight MobileNet-based convolutional neural network with EEG-derived features, combining a Squeeze-and-Excitation (SE) block and a Coordinate Attention (CA) model to capture inter-channel and temporal dependencies. A BiLSTM layer is employed to improve performance by modeling sequential patterns in both forward and backward directions. The results of the experiment demonstrate that the accuracy is $\mathbf{70.69 \%}$ in valence and $\mathbf{71.61 \%}$ in arousal on the DEAP dataset.