Performance comparison of different interpretative algorithms utilized to derive sleep parameters from wrist actigraphy data

ABSTRACT

We compared performance of four popular interpretative algorithms (IAs), i.e., Cole–Kripke, Rescored Cole–Kripke, Sadeh, and UCSD, utilized to derive sleep parameters from wrist actigraphy data. We conducted in-home sleep study of 40 healthy adults (17 female/23 male; age 26.7 ± 12.1 years), assessing sleep variables both by Motionlogger® Micro Watch Actigraphy (MMWA) and Zmachine® Insight+ electroencephalography (EEG). Data of MMWA were separately scored per 30 sec epochs by each of the four popular IAs, and data of the Zmachine were also scored per 30 sec epochs by its proprietary IA. In reference to the EEG Zmachine method, all four of the MMWA algorithms showed high (~94 to 98%) sensitivity and moderate (~42 to 54%) specificity in detecting Sleep epochs. All of them significantly underestimated Sleep Onset Latency (SOL: ~9 to 20 min), and all of them, except the Sadeh IA, significantly underestimated Wake After Sleep Onset (WASO: ~22 to 25 min) and overestimated Total Sleep Time (TST: ~32 to 45 min) and Sleep Efficiency (SE: ~7 to 9%). The Sadeh IA showed significantly smaller bias than the other three IAs in deriving WASO, TST, and SE. Overall, application of ‘Rescoring Rules’ improved performance of the Cole–Kripke IA. The Sadeh and Rescored Cole–Kripke IAs exhibited highest agreement with the EEG Zmachine method (Cohen’s Kappa: ~51%), while the UCSD IA exhibited lowest agreement (Cohen’s kappa: ~47%). However, minimum detectable change across all sleep parameters was smallest with use of the UCSD IA and, except for SOL, largest with use of the Sadeh algorithm. Findings of this study indicate the Sadeh IA is most appropriate for deriving sleep parameters of healthy adults, while the UCSD IA is most appropriate for evaluating change in sleep parameters over time or in response to medical intervention.

KEYWORDS:

• Actigraphy
• sleep
• Cole–Kripke algorithm
• Sadeh algorithm
• UCSD algorithm
• accuracy
• electroencephalography (EEG)
• epoch by epoch analysis
• rescoring rules

Disclosure statement

The authors have no conflicts of interest that influence the content of this study.

Additional information

Funding

This work was supported by the Robert and Prudie Leibrock Professorship in Engineering at the University of Texas at Austin.