Relational local electroencephalography representations for sleep scoring.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Computational sleep scoring using polysomnography (PSG) is clinically successful.
• Single electroencephalographic (EEG) channel models struggle with Rapid Eye Movement (REM) sleep scoring quality.
• Existing Long Short-Term Memory (LSTM) models inadequately represent distant EEG epochs.

Purpose of the Study:

• To develop a novel EEG representation model to improve single-channel sleep scoring.
• To address the limitations of LSTM models in capturing long-range temporal dependencies in EEG data.
• To enhance Rapid Eye Movement (REM) sleep scoring accuracy using only EEG signals.

Main Methods:

• Introduction of ENGELBERT (electroEncephaloGraphic Epoch Local Bidirectional Encoder Representations from Transformer), a Transformer-based model.
• Joint attention mechanism applied to multiple past and future EEG epochs.
• Local attention on overlapping windows to reduce computational complexity to linear.

Main Results:

• ENGELBERT achieved state-of-the-art macro F1-scores in single-EEG sleep scoring experiments.
• REM F1-scores were improved to at least 86%.
• The performance gap between single-EEG and PSG-based methods was reduced to less than 1 percentage point.

Conclusions:

• ENGELBERT effectively represents distant EEG epochs, overcoming LSTM limitations.
• The model enables versatile scoring from sub-one-hour to all-day periods.
• ENGELBERT significantly enhances single-EEG channel sleep scoring, approaching PSG-level accuracy, particularly for REM sleep.