Stacked EEG Spectrograms and an Attention-Augmented CNN-LSTM for Subject-Independent Emotion Recognition.

BACKGROUND: EEG provides direct neural measurements with high temporal resolution for emotion recognition. However, many spectrogram-based approaches process channels independently or integrate channel information only at later stages. NEW METHOD: We propose a stacked spectrogram representation that vertically concatenates channel-wise EEG spectrograms, preserving temporal, spectral, and channel-structured information at the input level. This representation is combined with an attention-augmented CNN-LSTM architecture for feature extraction, temporal modeling, and adaptive weighting of informative segments. RESULTS: Under LOSO cross-validation, the proposed framework achieved 82.34% accuracy and a macro-F1 score of 0.83 in four-class emotion classification. On SEED-IV, it achieved 84.51% accuracy and a macro-F1 score of 0.86. COMPARISON WITH EXISTING METHODS: The proposed method outperformed single-channel CNNs, stacked VGG16, and CNN-LSTM baselines, demonstrating competitive performance with lower computational complexity. CONCLUSIONS: Input-level channel integration effectively improves subject-independent EEG emotion recognition and provides a practical solution for consumer-grade EEG applications.