An interpretable frequency-band-aware framework for EEG-based emotion recognition using ensemble classifiers and CNN-LSTM networks

Accurate EEG-based emotion recognition remains challenging because emotional EEG patterns are highly subject dependent and vary across frequency bands and temporal segments. This study proposes an interpretable frequency-band-aware framework for binary valence and arousal recognition using DEAP as the primary dataset and DREAMER for additional cross-dataset evaluation. Preprocessed EEG signals were decomposed into Delta, Theta, Alpha, Beta, Gamma, and 0.5–45 Hz bands, segmented with 3-s windows under 0-, 1-, and 2-s overlap settings, and evaluated using leakage-free Leave-One-Subject-Out (LOSO) validation. Five EEG-informed feature representations: Differential Entropy (DE), Log Power Spectral Density (Log-PSD), Wavelet Entropy (WE), PLV-based functional connectivity and hemispheric asymmetry, were first compared using XGBoost, after which Differential Entropy was selected as the most stable feature. DE features were then assessed across frequency bands and overlap settings using XGBoost, Random Forest and CNN-LSTM. Gamma-band DE with 3-s windows and 2-s overlap provided the best performance, with CNN-LSTM achieving 82.10 ± 4.70% valence and 79.40 ± 5.20% arousal accuracy on DEAP. The results also showed that 2-s overlap improves temporal continuity and that Gamma and Beta bands carry stronger emotion-related information than lower-frequency or full-band features. SHAP analysis using a multi-band DE CNN-LSTM model identified Gamma DE as the most influential feature, supporting model interpretability. On DREAMER, the same configuration achieved 77.00 ± 5.00% valence and 74.00 ± 5.00% arousal accuracy, indicating consistent cross-dataset behavior.