Unsupervised Multitaper Spectral Method for Identifying REM Sleep in Intracranial EEG Recordings Lacking EOG/EMG...

Area of Science:

• Neuroscience
• Sleep Science
• Biomedical Engineering

Background:

• Large datasets of human intracranial EEG (iEEG) exist for clinical use.
• These recordings often lack simultaneous EOG/EMG, hindering sleep stage classification (Wake vs. REM).
• Existing iEEG-only methods struggle to differentiate Wake and REM sleep accurately.

Purpose of the Study:

• To develop an accurate method for classifying Wake vs. REM sleep using iEEG data alone.
• To enable the analysis of extensive, previously inaccessible iEEG datasets for sleep research.
• To improve the utilization of clinical iEEG recordings for understanding sleep architecture.

Main Methods:

• A novel unsupervised multitaper alpha-rhythm classifier was developed.
• The method generalizes K-means clustering for multitaper spectral eigencoefficients.
• Performance was assessed on eight subjects and compared to a classical power detector.

Main Results:

• The multitaper classifier identified 36±6 min of REM sleep per night.
• It mislabeled less than 10% of epochs for most subjects (human reliability ~80%).
• The proposed method outperformed the classical power detector.

Conclusions:

• The multitaper alpha-rhythm classifier accurately distinguishes Wake and REM sleep from iEEG alone.
• This method significantly increases the usability of clinical iEEG data for sleep studies.
• Accurate generalization is likely with single-night data, paving the way for large-scale iEEG sleep analysis.