Objective assessment of sleep regularity in 60 000 UK Biobank participants using an open-source package Free

Dear Editor,

Human health and behavior are regulated by a complex and extensive network of circadian clocks. These clocks are entrained by rhythmic signals in the environment, such as daily light exposure. In individuals who have irregular sleep schedules, these signals that furnish time-of-day information to the circadian system are less robust, which may cause circadian disruption and poor health outcomes [1]. Sleep regularity can be quantified using the Sleep Regularity Index (SRI) [2], a metric that compares sleep patterns between consecutive days (sleeping at similar times each day results in a high SRI). The SRI captures day-to-day variability in bedtime, waketime, sleep duration, naps, and awakenings during sleep [3]. Lower SRI has been associated with substantially increased risk for obesity, diabetes, cardiovascular disease, hypertension, and depressed mood [4–6]. Although sleep regularity is now recognized as a critical dimension of sleep health, there are barriers to measuring and reporting sleep regularity consistently. First, there are no open-source options for calculating sleep regularity, meaning it cannot be computed with the same ease as other common sleep metrics. Second, there is a lack of clear benchmarks for what represents a high or low level of sleep regularity at a population level, both for the SRI and other sleep regularity metrics [3, 4, 6]. We developed an open-source package for computing SRI from accelerometer data, and we applied it to the single largest accelerometer sample available to researchers, within the UK Biobank.

The UK Biobank is one of the largest and most comprehensive human health datasets [7]. In total, 103 104 participants (age [M ± SD] = 62.3 ± 7.9 years; 56.2% female) wore a wrist accelerometer (Axivity AX3 [8]) to record daily rest–activity patterns for 1 week, between June 2013 and January 2016. We used the accelerometer data to estimate daily sleep onset and offset times with a validated, widely used R package (“GGIR” [9]). Since it is not possible to correctly calculate SRI from GGIR output alone, we developed our own accompanying open-source R package. Our package calculated sleep–wake state at the epoch level and, unlike GGIR, it allowed multiple sleep bouts per day (allowing naps, fragmented sleep, or awakenings to be correctly factored into SRI calculation). The package also identified miscalculated sleep onset/offset times (5.7% of all nights in this dataset), which can lead to incorrect SRI scores (see Supplementary Material). After excluding participants with fewer than 5 days (120 h) of 24 h separated epoch pairs containing valid sleep–wake data, the SRI was calculated in 60 997 participants (age [M ± SD] = 62.7 ± 7.8 years; 55.1% female).

The SRI scores had a median of 81.0 and interquartile range of 73.8–86.3 (M ± SD = 78.8 ± 10.7), shown in Figure 1, A with a higher score indicating more regular sleep patterns. The distribution of SRI scores was nonnormal with negative skew (KS test, D(60 997) = 0.99, p < .0001), consistent with distributions reported in smaller samples [4], and 99% of individuals had SRI scores between 36.0 and 95.0. We observed a monotonic relationship between SRI and other common measures of sleep variability, including standard deviations in sleep onset, sleep offset, and sleep duration, shown in Figure 1, B. These findings enable SRI scores to be related to equivalent values for other sleep variability measures, facilitating comparisons between studies that have reported different measures (see Supplementary Material). Across the sample, a cutoff of SRI <70 was comparable to SD >1.9 h for variability in sleep onset and offset, and SD >1.6 h for sleep duration. One in five individuals had an SRI below 71.6 (irregular), and one in five had an SRI above 87.3 (regular; Figure 1, C).

Figure 1.

(A) Distribution of SRI scores in 60 997 UK Biobank participants, grouped by percentile ranges. Inset raster plots of sleep–wake patterns in three participants represent regular (SRI = 94; top), slightly irregular (SRI = 70; middle), and highly irregular (SRI = 34; bottom) sleep–wake patterns. (B) Mean intraindividual standard deviation in sleep timing and duration for each percentile range [color matched to (A)], calculated using each participant’s 1-week data collection period. (C) Comparative distributions of standard deviation in sleep timing and duration for upper (SRI >87.3) and lower (SRI <71.6) quintiles.

Open in new tabDownload slide

Sleep regularity was related to several self-reported demographic variables, collected between 2006 and 2010. Lower SRI scores were found in those who were male (difference: −1.2, t-test, p < .0001), had higher material deprivation (Townsend Deprivation Index, bivariate correlation, r(60 922) = −0.11, p < .0001), had lower yearly household income (ANOVA, p < .0001), or had lower-level educational qualifications (ANOVA, p < .0001) (see Supplementary Material). Those of white ethnicity had significantly higher SRI than all other ethnic groups (2.6–6.8 points higher), and black ethnicity significantly lower (2.2–6.8 points lower; Kruskal–Wallis, p < .0001). Above the age of 65, sleep regularity decreased with age (bivariate correlation, r(27 918) = −0.02, p = .002). There was no relationship between age and SRI in people younger than 65 (p = .47), or across the whole sample age range (p = .13). Shift workers exhibited significantly lower SRI scores than nonshift workers (M ± SD = 75.9 ± 12.0 vs. 79.3 ± 10.1, t-test, p < .0001). Employed people had significantly higher SRI than those who were Sick/Disabled, Unemployed, or Students (1.0–4.9 points higher), and lower SRI than Retired, Home/Family Caretakers, or Volunteers (0.2–1.6 points lower; Kruskal–Wallis, p < .0001). Together, these findings indicate that irregular sleep–wake patterns are associated with a complex set of individual and environmental factors, particularly socioeconomic disadvantage.

Sleep regularity is increasingly recognized as a fundamental determinant of health, and is a stronger predictor of cardiometabolic outcomes and quality of life than sleep duration [4, 10]. The norms we begin to establish here provide a reference for researchers and clinicians intending to quantify sleep regularity with the SRI. Relative to other measures of variability in sleep timing, the SRI offers two key advantages: (1) it compares sleep–wake patterns between consecutive days (i.e. on a circadian timescale), meaning it may assess circadian disruption [3]; (2) the SRI makes no assumptions about the structure of sleep (e.g. no assumption of one main sleep bout), making it applicable to populations with unusual sleep structure, such as shift workers or individuals with highly fragmented sleep [1, 11]. We have developed a package to calculate SRI scores from accelerometer or binary sleep–wake data in R, available at https://github.com/dpwindred/sleepreg. The package can be used in combination with GGIR or as a standalone method for calculating SRI from preprocessed sleep–wake data. Processing of accelerometer data is based on the methods used here—estimating sleep–wake timing using GGIR, identifying fragmented sleep patterns, and identifying and excluding nights containing probable estimation errors. Our package also extracts percentile of calculated SRI scores, allowing other investigators’ datasets to be compared to the UK Biobank SRI distribution, and generates raster plots of sleep–wake patterns. Given the widespread availability and use of consumer and research accelerometer data, our package will democratize the use of sleep regularity as an indicator of circadian health.

Funding

This research was carried out under UK Biobank Project ID 6818.

Disclosure Statement

MKR declares consultancy fees received from CGT Catapult. SWC and AJKP have received research funding from Versalux and Delos, and previously were investigators on projects funded by the Cooperative Research Centre for Alertness, Safety and Productivity. SWC has received research funding from Beacon Lighting, and has consulted for Dyson.

References