The potential of ensemble-based automated sleep staging on single-channel EEG signal from a wearable device
Machine-learning-based sleep staging models have achieved expert-level performance on standard polysomnographic (PSG) data. However, their application to EEG recorded by wearable devices remains limited by non-conventional referencing montage and the lack of benchmarking against PSG. Here, we tested whether an ensemble of state-of-the-art automatic staging algorithms can reliably classify sleep from a customized configuration of the ZMax headband, adapted to record a single fronto-mastoid EEG channel. A total of 35 nights of simultaneous ZMax and PSG recordings were acquired in a home setting, amounting to 250.02 hours of analysable data from 10 healthy participants. PSG data were scored according to AASM criteria by two independent experts from different sleep centres, with discrepancies resolved to obtain a consensus hypnogram. ZMax signal was processed using four machine-learning algorithms (YASA, U-Sleep, SleepTransformer, DeepResNet), whose predictions were further combined into a final ensemble scoring through soft-voting. The ensemble scoring achieved almost perfect agreement with human consensus staging (night-level mean {+/-} SD; accuracy = 88.83% {+/-} 2.84%, Cohens {kappa} = 84.10% {+/-} 4.52%, and Matthews Correlation Coefficient = 84.54% {+/-} 4.23%). It showed excellent predictive accuracy for REM (F1-score = 93.99%), N3 (89.53%), N2 (87.93%), and wakefulness (86.37%), with lower performance for N1 (53.20%). These findings support the deployment of an ensemble scoring approach based on state-of-the-art sleep staging algorithms on ultra-minimal, mastoid-referenced EEG setups. This paradigm opens the way to the integration of data from modern wearable technologies into traditional PSG-based sleep research, overcoming longstanding barriers to ecological and large-scale sleep monitoring.