An algorithm for actigraphy-based sleep/wake scoring: Comparison with polysomnography

Abstract

Objective: To evaluate the accuracy of actigraphy against polysomnography (PSG) as gold standard using a newly developed algorithm for sleep/wake discrimination that explicitly models the temporal structure of sleep.

Methods: PSG was recorded in 11 men and 9 women (age 71.1±5.0) to evaluate suspected neuropsychiatric sleep disturbances. Simultaneously, wrist actigraphy was recorded, from which 37 features were computed for each 1-min epoch. We compared prediction of PSG-derived sleep/wake states for each of these features between our newly developed algorithm, and four state-of-the-art algorithms. The algorithms were evaluated using a leave-one-subject out cross validation.

Results: The new algorithm classified 84.9% of sleep epochs (sensitivity) and 74.2% of wake epochs correctly (specificity), leading to a sleep/wake scoring accuracy of 79.0%. Four out of five sleep parameters were estimated more accurately by the new algorithm than by state-of-the-art algorithms.

Conclusion: The proposed algorithm achieved a significantly higher specificity than state-of-the-art algorithm, with only minor decrease in sensitivity for patients with sleep disorders. We assume this reflects the capability of the algorithm to explicitly model sleep architecture.

Significance: The unobtrusive assessment of sleep/wake cycles is particularly relevant for patients with neuropsychiatric diseases that are associated with sleep disturbances, such as depression or dementia.

Keywords: Actigraphy; Hidden Markov Model; Polysomnography; Sleep/wake scoring.