Sleep Analysis Using Longitudinal Ear-EEG Recordings

Ear-EEG offers significant advantages for longitudinal sleep studies since it is less intrusive and more user-friendly compared to traditional scalp EEG. The feasibility of longitudinal sleep studies further relies on automated sleep analysis algorithms. This study proposes a systematic and robust procedure for sleep analysis with a key focus on identifying recordings with poor data quality. The method is based on the USleep sleep scoring model and leverages the model's confidence score as a measure of signal quality. This strategy is based on the observation that there is a high correlation between the sleep model's confidence score and Cohen's kappa between manual and model annotations. Notably, the procedure does not rely on manually labeled data or other manual steps. The procedure was evaluated on a novel dataset comprising 574 sleep recordings from 24 chronic pain patients. The procedure distinguished recordings with kappa values above and below 0.6 with an accuracy of 91.9%. Importantly, the exclusion criteria did not systematically eliminate recordings with poor sleep quality metrics, such as low sleep efficiency or frequent sleep stage transitions. Furthermore, the study highlights the benefits of multiple-night sleep studies by visualizing the inter-night variability in each subject. In conclusion, the proposed procedure effectively excluded poor-quality recordings, enabling robust analysis of sleep patterns in patients.