Optimizing wearable single-channel electroencephalography sleep staging in a heterogeneous sleep-disordered...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Sleep Medicine

Background:

• Automated sleep staging using wearable electroencephalography (EEG) is limited by data availability in clinical populations.
• Transfer learning offers a strategy to overcome data limitations for sleep disorder research.

Purpose of the Study:

• To investigate transfer learning for optimizing single-channel EEG sleep staging in individuals with sleep disorders.
• To evaluate different model training strategies for wearable EEG devices.

Main Methods:

• Acquired 52 single-channel frontopolar headband EEG recordings alongside polysomnography (PSG).
• Compared three training strategies: pretraining, training-from-scratch, and fine-tuning.
• Evaluated performance using 10-fold cross-validation on headband recordings.

Main Results:

• Fine-tuning achieved the highest performance (κ = .778) for 5-stage sleep classification.
• Performance was significantly better than pretraining (κ = .769) and training-from-scratch (κ = .733).
• No significant differences or biases were observed in clinically relevant sleep parameters.

Conclusions:

• Deep transfer learning is crucial for optimizing performance with limited data in sleep staging.
• Conventional PSG and headband EEG data exhibit strong similarities.
• The headband, classification model, and fine-tuning methodology are viable for clinical sleep monitoring.