Performance evaluation of an automated single-channel sleep–wake detection algorithm

Abstract

Background

A need exists, from both a clinical and a research standpoint, for objective sleep measurement systems that are both easy to use and can accurately assess sleep and wake. This study evaluates the output of an automated sleep–wake detection algorithm (Z-ALG) used in the Zmachine (a portable, single-channel, electroencephalographic [EEG] acquisition and analysis system) against laboratory polysomnography (PSG) using a consensus of expert visual scorers.

Methods

Overnight laboratory PSG studies from 99 subjects (52 females/47 males, 18–60 years, median age 32.7 years), including both normal sleepers and those with a variety of sleep disorders, were assessed. PSG data obtained from the differential mastoids (A₁–A₂) were assessed by Z-ALG, which determines sleep versus wake every 30 seconds using low-frequency, intermediate-frequency, and high-frequency and time domain EEG features. PSG data were independently scored by two to four certified PSG technologists, using standard Rechtschaffen and Kales guidelines, and these score files were combined on an epoch-by-epoch basis, using a majority voting rule, to generate a single score file per subject to compare against the Z-ALG output. Both epoch-by-epoch and standard sleep indices (eg, total sleep time, sleep efficiency, latency to persistent sleep, and wake after sleep onset) were compared between the Z-ALG output and the technologist consensus score files.

Results

Overall, the sensitivity and specificity for detecting sleep using the Z-ALG as compared to the technologist consensus are 95.5% and 92.5%, respectively, across all subjects, and the positive predictive value and the negative predictive value for detecting sleep are 98.0% and 84.2%, respectively. Overall κ agreement is 0.85 (approaching the level of agreement observed among sleep technologists). These results persist when the sleep disorder subgroups are analyzed separately.

Conclusion

This study demonstrates that the Z-ALG automated sleep–wake detection algorithm, using the single A₁–A₂ EEG channel, has a level of accuracy that is similar to PSG technologists in the scoring of sleep and wake, thereby making it suitable for a variety of in-home monitoring applications, such as in conjunction with the Zmachine system.

Keywords:

• EEG
• sleep–wake detection
• algorithm
• Zmachine
• automatic sleep scoring
• single channel

Disclosure

The work was performed at, and supported by, General Sleep Corporation (a division of Consolidated Research, Inc., Euclid, OH, USA). Drs Kaplan and Wang are employees of General Sleep Corporation. Dr Loparo is affiliated with General Sleep Corporation and Nord Professor of Engineering in the Department of Electrical Engineering and Computer Science at Case Western Reserve University. Ms Kelly is a graduate student in the Department of Psychology at University of Arizona. Dr Bootzin is a Professor in the Department of Psychology at University of Arizona and a consultant to General Sleep Corporation. The authors report no other conflicts of interest in this work.