Hybrid EEG Feature Learning Method for Cross-Session Human Mental Attention State Classification.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Decoding mental attention states from electroencephalogram (EEG) is vital for applications like cognitive monitoring and brain-computer interfaces (BCIs).
• Existing EEG methods struggle with cross-session and inter-subject variability, limiting their real-world applicability.
• Robust attention decoding requires models that generalize beyond specific users or recording sessions.

Purpose of the Study:

• To develop a hybrid feature learning framework for robust classification of mental attention states (focused, unfocused, drowsy).
• To enhance the generalizability of EEG-based attention decoding across different sessions and individuals.
• To establish a foundation for practical, continuous mental state monitoring systems.

Main Methods:

• A unified pipeline integrating preprocessing, channel-wise spectral feature extraction (STFT), and connectivity features (functional and structural).
• A two-stage feature selection combining correlation-based filtering and random forest ranking for relevance and dimensionality reduction.
• Support Vector Machine (SVM) for efficient and generalizable final classification.

Main Results:

• Achieved high classification accuracies of 86.27% and 94.01% on two cross-session and inter-subject EEG datasets.
• Significantly outperformed traditional EEG-based attention decoding methods.
• Demonstrated the effectiveness of integrating connectivity-aware features with spectral analysis.

Conclusions:

• Integrating connectivity-aware features with spectral analysis significantly enhances the generalizability of attention decoding models.
• The proposed hybrid framework offers a promising approach for real-world EEG-based mental state monitoring.
• This work paves the way for more robust and adaptive brain-computer interfaces.