An Effective and Interpretable Sleep Stage Classification Approach Using Multi-Domain Electroencephalogram and...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Accurate sleep staging is essential for diagnosing sleep disorders and assessing sleep quality.
• Current deep learning methods for automated sleep staging offer limited real-world applicability due to complexity and lack of interpretability.
• Classical machine learning pipelines offer a potentially more accessible approach to automated sleep staging.

Purpose of the Study:

• To develop an effective and interpretable automated sleep staging scheme using a classical machine learning pipeline.
• To enhance sleep staging accuracy by integrating multi-domain EEG features with novel EOG features.
• To provide a transparent and practical solution for sleep stage classification.

Main Methods:

• Extracted multi-domain features from preprocessed electroencephalogram (EEG) signals.
• Developed novel electrooculogram (EOG) features to characterize sleep stages.
• Employed a two-step feature selection strategy (F-score pre-filtering and XGBoost ranking) followed by XGBoost classification.
• Validated the model using a double-cross-validation procedure on the Sleep-EDF dataset.

Main Results:

• Achieved competitive classification performance: 87.0% accuracy, 86.6% F1-score, and 0.81 Kappa coefficient on the Sleep-EDF dataset.
• Demonstrated comparable results to state-of-the-art deep learning methods.
• Provided interpretability through feature importance analysis, highlighting the model's transparency.

Conclusions:

• The proposed classical machine learning approach is effective for automated sleep staging.
• The model offers a practical alternative to complex deep learning methods due to its low complexity and interpretability.
• The scheme shows significant potential for real-world applications in sleep quality assessment and sleep disorder diagnosis.