Introduction

Circadian rhythms are endogenous cycles in physiological, hormonal, and behavioral processes largely synchronized to the external 24-hour light-dark cycle. The sleep-wake cycle is perhaps the most notable biological process that follows the 24-hour circadian rhythm.¹ The timing of light exposure is the primary external driver of circadian rhythms. Therefore, rapid changes in sleep timing can cause circadian misalignment through fluctuating light-dark exposure.²

Circadian misalignment is associated with several age-related diseases, including cancer and cardiovascular disease (CVD).^3–5 However, the health impacts of irregular sleep wake timing are still emerging. This remains an important area of study since modern societal and lifestyle trends, including exposure to artificial and blue light at night, longer work hours, shift work, and the 24-7 lifestyle have blurred the distinction between day and night, increasing the propensity for circadian disruption.⁶ The present study assessed the relationship between sleep regularity and the risk of incident all-cause mortality, cancer mortality, and CVD mortality in the UK Biobank (UKB). We measured sleep regularity via accelerometry to calculate the sleep regularity index (SRI), a new metric sensitive to differences in sleep-wake timing on a circadian timescale.

Methods

Participants

Over 500,000 adults aged 40 to 69 years were recruited to the UKB cohort between 2006 and 2010 across 22 assessment centers. Participants were invited by the UK National Health Service patient registers, resulting in a 5·5% participation rate. Respondents were more likely to be older, female, and less likely to live in socioeconomically deprived areas than the general population.⁷ Baseline demographics, medical history, lifestyle, vitals, and blood samples were collected. A total of 106,053 participants completed a 7-day wrist-worn accelerometer study through random selection between February 2013 and December 2015.

Measurement of sleep regularity

Accelerometry data were collected using a wrist-worn device (Axivity AX3, United Kingdom) over a 7-day/night period. Estimated sleep status (awake or asleep) at a given time was calculated using the open-source R package GGIR version 2·7-1,⁸ using available algorithms.^{9, 10} To distinguish sleep from sustained periods of inactivity without reference to a sleep diary (not available in the UKB), GGIR uses an algorithm to determine a daily ‘sleep period time window’ for each participant.¹¹ This defines the time window between the onset and end of the main daily sleep period, during which periods of sustained inactivity are interpreted as sleep. The algorithm does not, by default, detect bouts of sleep outside of this window and hence is not able to identify naps. Accelerometry data of low quality were removed using established UKB criteria (Appendix 1). Most participants (88%) provided complete accelerometry data. Participants with fewer than two valid SRI measurements (i.e., less than 2 contiguous 24-hour wear periods; <1%) were excluded. In total, 88,975 (84%) participants provided valid SRI data and were included in the study.

The SRI captures the probability of a participant being in the same state (asleep or awake) at any two time points 24LJhours apart.¹² An individual who sleeps and wakes at precisely the same time each day would have an index of 100, whereas an individual who sleeps and wakes at entirely random times would have an index of 0. Each participant provided k-1 SRI measurements (where k is the number of valid 24-hour periods), one for each contiguous two-day pair. These SRI measurements were averaged using a linear mixed effects model with a random intercept for the participant and fixed effects for the day of the week and daylight savings transition. The average SRI was standardized over the day of the week and daylight savings transitions, so all SRI results were comparable.

Mortality ascertainment

Mortality occurrence was identified through linkage with NHS Digital for participants from England and Wales and the NHS Central Register for participants from Scotland, with complete records available until January 2022. Death records included the date of death and the ICD-10 code for the primary cause. ICD-10 codes I00-I99 and C00-C97 defined CVD and cancer mortality, respectively.

Ascertainment of disease status at baseline

Past or prevalent cancer (ICD-10 codes D00-D09 and D37-D48), diabetes (codes E10-E14), mental and behavioral disorders (codes F00-F99), nervous system disorders (codes G00-G99), and CVD (codes I00-I99) at the time of the accelerometry study were ascertained through self-report at the UKB baseline session and through linkage with hospital inpatient records using the above ICD-10 codes. Linkage with hospital inpatient records was also used to identify disease occurrence between the UKB baseline session and the time of the accelerometry study.

Data Analysis

Data analysis was performed using R version 4·2·1. Cox proportional hazards models were used to examine associations between the SRI and incident all-cause mortality, CVD mortality, and cancer mortality. Surveillance for mortality commenced from the time of accelerometry (2013-2015) until the end of follow-up (January 2022), with a median follow-up time of 7·1 years (Q1, Q3: 6·6, 7·6). Non-events were censored at the last date they were known to be alive. For CVD and cancer mortality, deaths from competing causes were censored at the time of death. The SRI and all continuous confounders were modeled with restricted cubic splines with knots at the 10th, 50th, and 90th percentiles to allow for departures from linearity. Effect modification was assessed by adding product terms to Cox models. Missing data were infrequent (< 2%) for most confounder variables and were imputed (10 imputations) by predictive mean matching using the aregImpute function of R package Hmisc.¹³

In addition to Cox models, discrete-time hazards models, including an interaction between SRI and time (aggregated into 3-month intervals and modeled with a restricted cubic spline with knots at the 5^th, 35^th, 65^th, and 95^th percentiles), were fitted to relax the assumption of proportionality and allow hazard ratios (HRs) to vary over time.¹⁴ The SRI by time interaction in this model provided a test of proportionality (a small p value would indicate strong evidence against the proportional hazards assumption). Time-varying HRs were then displayed visually. In cases where HRs showed clear time-variation (i.e., hazards were non-proportional), we nonetheless present HRs from the Cox models as these can be interpreted as a weighted average of the time-varying HRs.¹⁵

The discrete-time hazards model for all-cause mortality was also used to estimate standardized cumulative incidence (risk) across levels of SRI, with confidence intervals obtained by bootstrapping.¹⁶ To reduce computation demand, only single imputation was used for the discrete-time hazards models.

All models were adjusted for the following variables that were selected using a directed acyclic graph (Appendix, Figure S1): age, sex, ethnicity (White, Asian, mixed race, Black, or other), Townsend deprivation index, retirement status (retired vs. all other work arrangements), shift work (shift worker vs. non-shift worker), sick or disabled (self-reported employment category), household income (ordinal with 5 levels), highest level of education (ordinal with 6 levels), smoking status (current, former, never), smoking (pack years), and use of sedative, antidepressant, or antipsychotic medication.

Sensitivity analyses

We fitted a second statistical model to determine whether the observed associations were independent of sleep time and disruption. Therefore, Model 2 included additional adjustments for overnight sleep duration and wake after sleep onset (WASO), averaged across accelerometry days (plus primary model covariates). In the second sensitivity analysis (Model 3), we adjusted for past or prevalent disease at baseline (cancer, CVD, mental and behavioural disorders, nervous system disorders, diabetes), in addition to the variables in the primary model. These variables were included as part of a sensitivity analysis as it is unclear whether they may be mediators or confounders of the SRI-mortality relationship. Long-standing irregular sleep may lead to prevalent disease (or a history of disease) at baseline and influence disease risk factors,^17–20 indicating that these disease variables may play a mediating role (and consequently should not be adjusted). Conversely, past or prevalent disease may have effects disruptive to regular sleep and these variables may therefore confound the SRI-mortality relationship. Disease risk factor variables body mass index (BMI), moderate and vigorous physical activity (accelerometry-derived), systolic blood pressure (BP), and use of BP lowering medication, in addition to the variables in Model 3, were included in a final sensitivity analysis (Model 4), as it is similarly unclear whether they may confound or mediate the SRI-mortality relationship.

Comparison of SRI with other regularity measures

Preliminary reports which identified irregular sleep as a potential CVD risk factor measured sleep regularity as the amount of deviation in sleep patterns from an individual’s average (i.e., the standard deviation [SD] of nocturnal sleep duration and sleep onset time).^{21, 22} To contrast these SD-based metrics with the SRI, we fitted independent Cox models (each with primary model covariates) and estimated HRs for all-cause mortality for each of the three measures. Additionally, we added the SRI to a model containing both SD-based regularity measures (alongside primary model covariates) to test whether the SRI contained additional mortality risk information beyond that captured by the two SD metrics.

Results

Table 1 displays sample characteristics. The final sample size was 88,975. There were 3010 all cause deaths during a median follow-up of 7·1 years (Q1, Q3: 6·6, 7·6). The most common primary cause of death was cancer (n = 1701, 57%) followed by CVD (n = 616, 20%).

Table 1:

SRI and all-cause mortality

We identified a non-linear association between the SRI and all-cause mortality hazard (p [global test of spline term] < 0·001) (Figure 1). Compared to the sample median (SRI = 61), mortality rates were highest among those with the most irregular sleep and decreased almost linearly as SRI approached its median, after which the decrease began to plateau (Figure 1). HRs, relative to the median SRI, were 1·53 (95% CI: 1·41, 1·66) for participants with SRI at the 5^th percentile (SRI = 41) and 0.90 (95% CI: 0·81, 1·00) for those with SRI at the 95^th percentile (SRI = 75), respectively. Standardized cumulative incidence curves for all-cause mortality are displayed for the SRI at the 5^th percentile, median, and 95^th percentile in Figure 2. There was little indication that hazard ratios varied according to age (p interaction = 0.48), sex (p = 0.36), household income (p = 0.62), sleep duration (p = 0.47), moderate to vigorous physical activity (p = 0.13), past or prevalent CVD (p = 0.48), or past or prevalent cancer (p = 0.29).

Figure 1.

Figure 2.

There was strong evidence against the proportionality assumption in the discrete-time hazards model (p [time x SRI interaction] < 0·001). Time-varying HRs for the 5^th and 95^th SRI percentiles compared to the median are displayed in the appendix (Figure S2). For the 5^th percentile relative to the median, HRs were greatest in the earliest period of follow-up (HRs> 2), declining until approximately 2·5 years, after which they remained approximately stable with a HR of around 1·5. There was no clear time variation in the HR for the 95^th percentile of SRI vs. the median.

CVD-specific mortality

The SRI was associated with CVD-specific mortality in the primary model (p [global] < 0·001; Figure 1). HRs, relative to the median SRI, were 1·88 (95% CI: 1·61, 2·21) and 0·93 (95% CI: 0·73, 1·20) for the 5^th and 95^th percentiles, respectively. There was no evidence of non-proportional hazards in the discrete-time hazards model (p [time x SRI interaction] = 0·57). There was little indication that HRs varied according to age (p interaction = 0.17), household income (p = 0.30), sleep duration (p = 0.69), moderate to vigorous physical activity (p = 0.95), past or prevalent CVD (p = 0.16), or past or prevalent cancer (p = 0.24). HRs for low SRI were larger for males than females (p interaction = 0.006; Figure S4).

Cancer-specific mortality

The SRI was associated with cancer mortality in the primary model (p [global] < 0·001). HRs, relative to the median SRI, were 1·36 (95% CI: 1·22, 1·53) and 0·89 (95% CI: 0·77, 1·02) for the 5^th and 95^th percentiles, respectively. There was strong evidence of non-proportional hazards in the cancer-mortality discrete-time hazards model (p [time x SRI interaction] < 0·001). HRs, for the 5^th percentile vs. the median, were large at the beginning of follow-up (HRs> 2) and declined until approximately four years, after which they were small (∼ 1·05) and compatible with the null (Appendix, Figure S3). There was no indication that HRs for the 95^th SRI percentile relative to the median varied over follow-up. There was little indication that hazard ratios varied according to age (p interaction = 0.48), sex (p = 0.36), household income (p = 0.62), sleep duration (p = 0.49), moderate to vigorous physical activity (p = 0.42), past or prevalent CVD (p = 0.82), or past or prevalent cancer (p = 0.12).

Figure 3.

Sensitivity analyses

Sensitivity analyses are displayed in the appendix (Figures S5-7). Overall, results were similar and not meaningfully altered following adjustments for sleep time and WASO (Model 2) or past or prevalent baseline disease, including cancer and CVD (Model 3). The SRI remained associated with mortality after further adjustments for past or prevalent diseases, BMI, systolic BP, BP treatment, and physical activity (Model 4), though effect sizes were attenuated. For example, when comparing the 5^th percentile to the median, HRs were 1·22 (95% CI: 1·07, 1·39) for all-cause, 1·43 (95% CI: 1·21, 1·69) for CVD, and 1·15 (95% CI: 1·01, 1·29) for cancer mortality.

Comparison of SRI with sleep duration SD and sleep onset time SD

The SRI was modestly negatively correlated with the sleep duration SD (-0.32) and sleep onset time SD (-0.42; see correlation matrix in Table S1). Figure 3 displays HRs, relative to the median, for the SRI, sleep duration SD, and sleep onset time SD. For each measure, greater sleep irregularity (i.e., lower SRI or higher SD representing more day-to-day variability) was associated with an increased all-cause mortality rate in independent models (all p [global] < 0·001). HRs, for low regularity compared to the median, were largest for the SRI (Figure 3). The addition of the SRI to a model containing both SD metrics (alongside primary model covariates) improved model fit (p [likelihood ratio test] < 0·001). Conversely, the addition of sleep duration SD and sleep onset time SD to a model containing the SRI (and primary model covariates) did not meaningfully improve model fit (p [likelihood ratio test] = 0·10).

Discussion

Among 88,975 individuals followed for a median of 7·1 years, there was a non-linear association between sleep regularity and the risk of mortality; mortality rates were highest in persons with the most irregular sleep and decreased approximately linearly as sleep regularity approached its median, after which the decrease began to plateau. Our findings were independent of past or prevalent illness (including cancer and CVD at baseline), sleep duration, sleep fragmentation, and other confounding factors. Overall, these data indicate a relationship between sleep regularity and longevity in a large community-based cohort.

Physiological processes associated with CVD and cancer are under circadian control. Mutations or deletions to circadian clock genes such as CLOCK, PER, and BMAL1 influence BP, endothelial function, and glucose homeostasis.^23–26 Both major (e.g., chronic shift work) and minor (e.g., daylight savings transitions) stressors to the circadian system have been associated with a higher risk of CVD.^{3, 5} Similarly, circadian misalignment has also been implicated in the pathogenesis of cancer. For example, circadian clocks are critical to the orchestration of cell division²⁷, and altered clock function can precipitate aberrant cell proliferation²⁸ as well as growth and DNA damage in cancer cells.²⁹ Many systems are under circadian influence, including the sleep-wake cycle, and less clear has been the extent to which differences in sleep regularity are related to negative health outcomes. We extend this research by demonstrating that differences in sleep regularity are associated with the risk of mortality from both CVD and cancer.

Whereas sleep regularity has not been examined with respect to incident cancer or mortality, the current findings extend research showing that greater sleep-wake variability, as measured by the SD of sleep onset or duration, was independently associated with a higher risk of incident CVD in the multi-ethnic study of atherosclerosis but not the UK Biobank.^{21, 22} We demonstrate that the SRI contains information about mortality risk beyond that contained in the SD of sleep duration and onset, whereas the converse was not the case. The SRI may be superior to the SD-based metrics because the SRI captures rapid changes in sleep patterns across consecutive days, as compared to the SD-based metrics which only calculate deviation from an individual’s average. Rapid changes in sleep timing have been hypothesised as being principally challenging for the circadian system to accommodate¹² which may, in turn, produce negative health outcomes.

We found evidence that hazard rates across levels of SRI were non-proportional (i.e., varied across the follow-up period) for all-cause and cancer mortality (which accounted for most deaths), though not for CVD mortality. For cancer mortality, HRs for low SRI compared to the median were largest in the earliest follow-up period and decreased thereafter. One plausible interpretation of this finding is that irregular sleep may be a manifestation of the underlying physiological processes of cancer itself or of cancer treatment (i.e., the SRI-cancer mortality association may be due to reverse causation). However, this thesis is challenged by the fact that associations between the SRI and cancer mortality remained similar after adjusting for past or prevalent cancer at baseline. In the case of CVD mortality, no such evidence of a decline in HRs over follow-up time was evident; a potential causal role of irregular sleep on CVD death cannot be easily ruled out.

Sleep of insufficient or excessive duration is associated with many adverse health outcomes, including increased mortality risk.³⁰ As of 2022, sleep duration was included by the American Heart Association in their Essential Eight guidelines for CVD prevention.³¹ However, sleep is far more complex than its habitual duration and quality, with sleep regularity receiving comparatively little attention. As sleep-tracking wearables become more accessible, objective measurement of sleep regularity has the potential for public and clinical use. Much like sleep duration, replicating the current findings across different samples will be necessary for establishing population norms and clinical targets. Furthermore, identifying the determinants of poor sleep regularity may be of import, not only considering biological factors, but broader social determinants that impact circadian rhythmicity (e.g., racial/ethnic disparities³², neighbourhood factors³³) and consequently overall health.

Our study is not without limitations. Firstly, the study was observational. We are, therefore, unable to establish cause and effect. Although we performed extensive analyses to control for confounding, we cannot exclude the possibility that our results are explained by residual confounding. As such, although therapies exist for improving sleep regularity, it’s not clear if these interventions are able to extend the lifespan. Second, sleep and wake were estimated through activity patterns from accelerometry. As compared to polysomnography, there is the potential to misclassify sleep and wake, although accelerometry is more suited to estimate circadian patterns over several days; there are several strengths to using accelerometry (e.g., days of continuous recording, minimal technical apparatus affecting sleep quality), making it the recommended clinical tool for assessing circadian rhythms.³⁴ In addition, sleep diaries in the UKB were not available. Consequently, the algorithm we used to determine sleep and wake relied on the identification of a main ‘sleep period time window’ and did not identify napping.

Circadian rhythms have a major influence on health and disease. Although sleep wake timing is under circadian control, research on sleep regularity as a risk factor for mortality was equivocal. These data suggest sleep regularity as an important correlate of longevity, independent of sleep duration, fragmentation, and quality.

Future work is needed to determine the underlying mechanisms to inform possible interventions to extend the lifespan.

Funding

Dr. Pase is supported by a National Health and Medical Research Council of Australia Investigator Grant (GTN2009264) with sleep research funding from the National Health and Medical Research Council of Australia (GTN1158384), National Institute on Aging (R01 AG062531) and Alzheimer’s Association (2018-AARG-591358). Dr. Baril is funded by the Banting Fellowship Program (#454104). The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Data Availability

Data from the UK Biobank are available, pending application approval from: https://www.ukbiobank.ac.uk/

Acknowledgements

This research has been conducted using the UK Biobank Resource under project ID 70607.

Data sharing

Data from the UK Biobank are available, pending application approval from: https://www.ukbiobank.ac.uk/

Appendix 1: Methods. Removal of low-quality accelerometer data

Accelerometry data of low quality were removed using established UKB criteria;

incongruity of self-reported wear time and accelerometer wear time data (5%); insufficient wear time (< 72 hours; 5%); and poorly calibrated data (<1%). Lastly, data were removed for participants in which GGIR was unable to determine a sleep window (5%) and for participants providing less than two valid SRI measurements (i.e., 2 24-hour wear periods; <1%). In total, 88,975 (84%) participants provided valid sleep regularity index data and were included in the study.

Appendix 2: STROBE Statement—Checklist of items that should be included in reports of cohort studies Item

Table S1:

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Figure S7.