| Summary: | There is a well-established association between sleep and health status, but the current gold-standard for analysing sleep, polysomnography, is too disruptive and expensive to enable longitudinal monitoring. There is, therefore, a growing interest in automated sleep scoring, or staging, using a combination of wearable technology to acquire cardio-respiratory vital signs and machine learning to learn how these vital signs vary with sleep state. However, sleep and the associated cardio-respiratory signals also change significantly with age, in part because of age-related changes in the autonomic nervous system, and this impacts the accuracy of wearable sleep staging methods. This paper investigates how the accuracy of a deep neural network model trained on the Sleep Heart Health Study database varies with the age of the subject. We show that the classification accuracy for each sleep stage decreases with age. We also present proof-of-concept analysis of longitudinal sleep data from a COVID-19 Challenge Study with a younger cohort (18 - 29 years of age), discuss the impact of having trained the deep neural network model on a database with an age range from 40 to 89+, and suggest how this issue may be addressed.Clinical relevance— This paper highlights how changes in sleep behaviour with age can affect neural network sleep staging using cardio-respiratory vital signs and machine learning, resulting in less accurate sleep staging in some age groups, and discusses potential methods for addressing this.
|
|---|