0445 A Novel Algorithm for the Estimation of Sleep States Based on Breathing and Movement

Abstract

Introduction We tested the diagnostic accuracy of the novel Nox BodySleep™ algorithm (Nox Medical, Iceland) for the estimation of sleep states from polygraphy (PG) sleep recordings based on features extracted from actigraphy and respiratory inductance plethysmography (RIP) belts. The algorithm automatically classifies epochs into three states, Wake, REM sleep and NonREM sleep. Validation was performed against polysomnography (PSG) in a sleep laboratory collective including patients with sleep disordered breathing (SBAS) and sleep related movements disorders. Methods Patients received PSG according to clinical routine. The recording was evaluated by the novel algorithm and the results were evaluated by descriptive statistics methods (IBM SPSS Statistics 25.0). Results We found a good Spearman correlation (r=0.8) and a bias of 11 minutes for the estimation of Total Sleep Time. Sleep Efficiency was also valued with a good Spearman correlation (r=0.7) and a bias of 1.6%. Wake phases were estimated with a F1 score of 0.64 while REM and Non-REM phases were evaluated with a F1 score of 0.73 and 0.82, respectively. Additionally, an overall accuracy of 0.8 and a Cohens kappa of 0.7 were found. Patients with sleep related movement disorders showed a slighly weaker correlation as patients with SBAS. Conclusion The algorithm shows a good diagnostic accuracy for the estimation of sleep states and significant sleep parameters. After validation on a larger patient collective, it could be used in the ambulatory and telemedical field to allow investigations comparable to the accuracy of a PSG. Support No support.

A systematic review of the performance of actigraphy in measuring sleep stages

“…The population characteristics are listed in Table 1. The majority of studies used data from healthy participants [15,16,17,18,19,20,21], whereas one study assessed the sleep staging performance in participants with sleep disorders [22]. The number of participants included in each study tended to be small.…”

“…Heuristic-based methods were not used by any study. Statistical learning methods were the predominant method class, including logistic regression [16], Hidden Markov Model [17], and tree-based methods [22,20]. Deep learning techniques have also been used in some studies [22,21].…”

“…Method development studies used N-fold/leave-one-out cross-validation [16,17,22,20,21]. Only one method development study reported the performance on datasets that the algorithm has not seen [21].…”

“…Even though five sleep stages are defined by AASM, because of the difficulty in the sleep stage identification from actigraphy, several sleep stages are combined into one to improve the agreement with PSG. For example, Walch et al[16] conducted the evaluation in a three-class setting (W/N1+N2/N3+R). While sleep stage grouping will help to obtain better performance, five-class staging performance needs to be reported to aid the comparison across different studies.…”