SSF-SET: A Discrete EEG Token-Based Framework for Sleep Stage Forecasting

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automated sleep staging using electroencephalogram (EEG) signals is crucial for health monitoring.
• Existing methods analyze past events, limiting their effectiveness for real-time, personalized sleep interventions.
• Predicting future sleep stages is essential for proactive sleep management.

Purpose of the Study:

• To develop a novel framework, the sleep stage forecaster with sleep EEG tokenizer (SSF-SET), for accurate prediction of future sleep stages.
• To enable personalized sleep interventions by forecasting sleep transitions using only past EEG data.
• To improve sleep quality through early detection of disruptive sleep stage changes.

Main Methods:

• The SSF-SET framework utilizes a sleep EEG tokenizer (SET) with a multi-branch transformer and LSTM encoder-decoder for feature extraction and quantization into informative tokens.
• A decoder-only transformer (SSF) is pre-trained for next-token prediction and fine-tuned using reinforcement learning with sequence-level rewards.
• The model predicts future sleep stages autoregressively without access to future EEG data during inference.

Main Results:

• SSF-SET demonstrated superior performance in predicting future sleep stages compared to direct forecasting methods on the SleepEDF20 and SleepEDF78 datasets.
• Achieved an accuracy of 0.596 and macro-F1 score of 0.516 on SleepEDF20.
• Attained an accuracy of 0.611 and macro-F1 score of 0.537 on SleepEDF78, confirming the effectiveness of quantized EEG tokens for autoregressive prediction.

Conclusions:

• Quantized sleep EEG tokens are effective for autoregressive prediction, enabling accurate forecasting of future sleep stages without future EEG data.
• The SSF-SET framework represents a significant advancement for closed-loop, personalized sleep interventions.
• This technology holds the potential to proactively improve sleep quality by anticipating and mitigating disruptive sleep transitions.