Extracting continuous sleep depth from EEG data without machine learning.

Area of Science:

• Neuroscience
• Sleep Science
• Signal Processing

Background:

• Human sleep is categorized into discrete stages using electroencephalography (EEG) and other biosignals.
• The ability to identify these human-defined stages using unsupervised machine learning methods remains unclear.
• Minimal pre-processing and generic analysis techniques are desired for broader applicability.

Purpose of the Study:

• To investigate if unsupervised methods can rediscover discrete human sleep stages from EEG data.
• To quantify the separability of sleep stages using the General Discrimination Value.
• To explore the potential of Principal Component Analysis (PCA) for identifying continuous sleep dynamics.

Main Methods:

• Analysis of overnight electroencephalography (EEG) data from sleeping human subjects.
• Transformation of time-domain EEG signals into the frequency domain for each 30-second epoch.
• Application of Principal Component Analysis (PCA) to epoch-wise frequency spectra to identify separable components.

Main Results:

• Raw and frequency-transformed EEG data showed minimal clustering of sleep stages.
• Principal Component Analysis (PCA) revealed significant separation of sleep stages in a low-dimensional subspace.
• The principal component C1(t) emerged as a robust, continuous variable correlating strongly with sleep depth and the hypnogram.
• C1(t) exhibited persistent trends during stable sleep stages, suggesting sleep as a continuum.

Conclusions:

• Unsupervised analysis of EEG data can identify a continuous measure of sleep depth (C1(t)).
• The component C1(t) offers a potential 'master variable' for understanding sleep dynamics and depth.
• Findings suggest sleep may be better conceptualized as a continuum rather than discrete stages.
• The properties of C1(t) could be leveraged for developing low-cost, single-channel sleep tracking devices.