Single-channel EEG-based sleep stage classification via hybrid data distillation.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Neuroscience

Background:

• Deep learning frameworks for automatic sleep stage classification require large electroencephalogram (EEG) datasets.
• Training on extensive EEG datasets presents significant computational challenges.
• The use of sensitive EEG data raises patient privacy concerns.

Purpose of the Study:

• To propose a hybrid data distillation method for single-channel EEG sleep stage classification.
• To enable training with reduced computational cost and enhanced privacy protection.
• To distill large, real-world EEG datasets into a small, privacy-preserving synthetic dataset.

Main Methods:

• Utilized gradient matching to optimize a randomly initialized synthetic dataset, rapidly minimizing performance gaps.
• Employed distribution matching to refine the synthetic dataset, ensuring global data consistency and avoiding training oscillations.
• Incorporated a novel mini-batch iteration technique to improve the synthetic dataset's ability to learn temporal dependencies.

Main Results:

• The proposed framework was validated on three public datasets, demonstrating robust performance.
• Achieved effective sleep stage classification using a distilled, synthetic dataset.
• The method proved efficient in training deep learning models for EEG analysis.

Conclusions:

• The study presents an efficient and robust hybrid data distillation algorithm.
• Offers a feasible solution for privacy-preserving sleep stage classification.
• Facilitates the development of deep learning models for sleep analysis with reduced data requirements and privacy risks.