Event-Aligned Deep Learning for EEG-Based Emotion Recognition

Accurate recognition of human emotional states from electroencephalography (EEG) remains challenging due to noise, inter-subject variability, and imprecise temporal annotation. This work presents an event-aligned EEG preprocessing and segmentation framework using the ECSMP dataset, where machine-generated markers enable precise extraction of emotionspecific intervals. Prefrontal EEG signals from eight channels are segmented into overlapping 4-second windows, producing a deep-learning-ready dataset for six-class emotion recognition. A comparative analysis is conducted between a hybrid CNN-BiLSTM model and a BiLSTM architecture to examine the impact of spatial-temporal feature learning. Under Leave-One-Subject-Out evaluation, the BiLSTM model achieves 45.24% accuracy (mean cross-subject accuracy $0.4514 \pm 0.0537$), closely aligning with the previously observed 45.65% benchmark. Conversely, the hybrid CNN-BiLSTM evaluated using 10-fold crossvalidation attains $\mathbf{5 5. 2 4 \%}$ accuracy (mean $0.5514 \pm 0.0537$), consistent with the earlier $\mathbf{5 4. 0 9 \%}$ result and a macro ROC-AUC of 0.838. These findings indicate that hybrid spatial-temporal modeling benefits from less restrictive validation settings, while purely temporal architectures demonstrate competitive robustness under subject-independent protocols. The proposed framework provides a reproducible benchmark for EEG-based emotion recognition with standardized comparative evaluation.