Domain Adaptation Using Large Scale Databases for Sleep Staging from a Novel In-Ear Sensor

Abstract

Background: Sleep disorders and poor sleep quality contribute to serious health problems and loss of quality of life, however there are few ways to reliably monitor the progression or treatment of sleep disorders for periods of more than a few nights. Long-term at-home monitoring of sleep quality is typically conducted via wrist-worn actigraphy sensors or subjective surveys, however such methods are far less reliable than full in-hospital polysomnography. Recent advances however have been made in using ear EEG as an alternative to scalp EEG for sleep staging , among other tasks. NextSense Inc. has developed a commercial EEG sensor which fits within the ear canal and which has been shown to be effective for seizure detection. Methods: We developed and evaluated the effectiveness of a deep transfer learning-based automated sleep staging algorithm for use with the NextSense device. A deep neural network was pre-trained on 1100 in-hospital polysomno-grams from the Wisconsin Sleep Cohort before fine-tuning on 8 recordings taken using the NextSense device. Results: Our approach achieved leave-one-subject-out cross-validation accuracy of 77.3% and Cohen’s κ of .647, beating the performance of multiple FDA-approved wearable scalp EEG sensors. Conclusion: These findings validate the use of the NextSense earbuds for use in sleep staging and open the possibility for at-home sleep monitoring using methods which are both more reliable than wrist-worn sensors and less cumbersome than bulky scalp EEG sensors.