Related Experiment Videos
Adaptive frequency band attention-guided CNN-BiLSTM for Spatial-Spectral-Temporal EEG emotion recognition
Rabita Hasan1, Sheikh Md Rabiul Islam2
1Department of Electronics and Communication Engineering (ECE), Khulna University of Engineering and Technology, Khulna, 9203, Bangladesh. rabitahasan32@gmail.com.
Brain Informatics
|July 21, 2026
Summary
This study introduces an interpretable Adaptive Frequency Band Attention (AFBA) framework for advanced electroencephalography (EEG)-based emotion recognition. Increased temporal overlap in EEG data significantly enhances recognition accuracy across various frequency bands and emotion dimensions.
Area of Science:
- Neuroscience
- Affective Computing
- Machine Learning
Background:
- Electroencephalography (EEG) offers high temporal resolution for emotion recognition but faces challenges due to signal complexity and subject variability.
- Existing methods often neglect temporal overlap effects, treat frequency bands uniformly, and lack spectral interpretability.
Purpose of the Study:
- To develop an interpretable framework for spatial-spectral-temporal EEG emotion recognition.
- To address limitations in existing methods by incorporating adaptive frequency band attention and evaluating temporal overlap effects.
Main Methods:
- A novel Adaptive Frequency Band Attention (AFBA)-guided multi-band 2D CNN-BiLSTM framework was proposed.
- EEG signals were decomposed into frequency bands (delta, theta, alpha, beta, gamma), and Power Spectral Density (PSD) features were extracted.
- The framework utilized 2D CNN for spatial-spectral features and BiLSTM for temporal dependencies, evaluating 3-second EEG windows with varying temporal overlap (0s, 1s, 2s).
Main Results:
- Increased temporal overlap in EEG data significantly improved emotion recognition performance across all frequency bands and emotion dimensions.
- With a 2-second overlap, the model achieved high accuracy on the DEAP dataset (e.g., 97.2% valence, 98.1% arousal) and SEED dataset (94.1% three-class accuracy).
- Ablation and SHAP analyses confirmed the importance of temporal modeling, AFBA, and high-frequency band activity (beta/gamma).
Conclusions:
- The proposed AFBA-guided framework enhances EEG-based emotion recognition by adaptively focusing on relevant frequency bands and leveraging temporal information.
- Optimizing temporal overlap in EEG signal processing is crucial for improving the accuracy and reliability of emotion recognition systems.
- The framework's interpretability and strong performance highlight its potential for real-world affective computing applications.