Benefits of Different Strategies to Adapt Sleep Scoring Models from Scalp- to Ear-EEG

Polysomnography is an extensive evaluation of several physiological measures and the gold standard technique for clinical sleep assessment. However, this technique is resource-expensive and often unfeasible over multiple nights. Ongoing research has shown that ear-EEG technology combined with the model-based automatic sleep scoring can be used for long-term sleep monitoring. More work is needed to robustly adapt clinical polysomnography-based sleep scoring models to at-home ear-EEG-based sleep patterns. Here, we investigated the main and combined benefits of utilizing three different strategies to adapt sleep scoring models from scalp- (part of polysomnography) to ear-EEG: 1) fine-tuning of the sleep scoring model to left-right mastoid scalp-EEG, 2) fine-tuning of the sleep scoring model to ear-EEG and 3) ensemble prediction. The results showed that all strategies applied in isolation improve the sleep scoring performance on ear-EEG data, relative to the not adapted model. With combined two or three strategies, sleep scoring performance on ear-EEG reached performance comparable with sleep scoring on scalp-EEG (from κ 0.71 to 0.83; from κ 0.67 to 0.77; from κ 0.57 to 0.68 in three data sets).Clinical Relevance- This study demonstrates that current crosshead ear-EEG technology, combined with advanced sleep scoring models, enables accurate at-home sleep monitoring. Furthermore, tailoring sleep scoring models to ear-EEG data enhances identification of the sleep architecture.