The Potential of Ensemble-Based Automated Sleep Staging on Single-Channel EEG Signal From a Wearable Device

Area of Science:

• Neuroscience
• Biomedical Engineering
• Sleep Medicine

Background:

• Machine-learning models show expert-level performance for sleep staging using standard polysomnography (PSG).
• Wearable device EEG presents challenges due to non-conventional referencing and lack of PSG benchmarking, limiting clinical application.

Purpose of the Study:

• To evaluate an ensemble of state-of-the-art sleep staging algorithms for reliable sleep classification using a customized ZMax headband (fronto-mastoid EEG channel).
• To benchmark wearable EEG sleep staging against traditional PSG by analyzing microstructural features and overall classification accuracy.

Main Methods:

• Simultaneous ZMax headband and PSG recordings were collected from 10 healthy participants (35 nights, 250.02 hours).
• Four machine-learning algorithms (YASA, U-Sleep, SleepTransformer, DeepResNet) processed ZMax EEG data.
• An ensemble scoring approach combined algorithm predictions via soft voting, with consensus hypnograms from two independent experts serving as the gold standard.

Main Results:

• The ensemble scoring demonstrated near-perfect agreement with human consensus staging (accuracy: 88.83%, Cohen's κ: 84.10%, MCC: 84.54%).
• High classification efficiency was observed for REM (F1: 93.99%), N3 (89.53%), N2 (87.93%), and wakefulness (86.37%), with lower performance for N1 (53.20%).
• Microstructural analysis confirmed strong correspondence between ZMax and PSG signals, indicating preservation of key sleep features.

Conclusions:

• An ensemble scoring approach using state-of-the-art algorithms is effective for ultra-minimal EEG setups like the ZMax headband.
• This validates the integration of wearable technology data into sleep research, overcoming barriers for ecological, large-scale sleep monitoring.