Association Between Accelerometer-Measured Irregular Sleep Duration and Type 2 Diabetes Risk: A Prospective Cohort Study in the UK Biobank

With more than 537 million prevalent cases and 6.7 million annual attributable deaths, type 2 diabetes is a rapidly growing but potentially preventable chronic disease worldwide (1). Identifying novel, modifiable diabetes risk factors could help in developing new interventions that reduce diabetes burden. The circadian timing system, consisting of the central brain clock and peripheral clocks in organs/tissues, regulates daily rhythmicity of glucose tolerance and insulin sensitivity (2). Misalignment between the circadian timing system and sleep-wake cycles due to behavioral, environmental, or genetic factors could result in insulin resistance and diabetes development (2).

Sleep duration, sleep quality, and other sleep disturbances (e.g., related to shift work) have been linked to diabetes risk in prior studies (3–5). Emerging evidence suggests that sleep irregularity/variability, which characterizes daily variations in sleep duration or timing, may be a novel risk factor for metabolic health independent of sleep quantity and quality. Irregular sleep may be a marker of habitual exposure to circadian disruption due to modern lifestyle (e.g., increased night activities) and environmental factors (e.g., noise, light at night). Irregular sleep may shift circadian phase from one night to another and disrupt circadian rhythms (6), leading to glucose dysregulation (7), metabolic syndrome (8), and, over time, an increased risk of insulin resistance and diabetes. There are multiple ways of measuring irregular sleep (9). Operationalized as within-person SD of sleep measures across multiple consecutive nights (10), prior cross-sectional studies have linked less regular sleep with metabolic abnormalities and diabetes (8,10). However, the only prospective cohort study to date found no significant association between sleep irregularity and incident diabetes in 2,107 Hispanics/Latinos, using the SD or the sleep regularity index (10,11). Given the modest sample size of this study, larger prospective studies are needed to determine whether irregular sleep is associated with risk for incident diabetes (10) and whether this association varies by demographic or other sleep-related factors.

It remains unknown whether irregular sleep duration, a potentially modifiable behavioral factor, interacts with genetic susceptibility to influence diabetes risk. Given the associations of irregular sleep duration with key diabetes precursors including obesity, inflammation, and metabolic syndrome, it is possible that, because of shared genetic factors, irregular sleep would be more strongly linked to incident diabetes risk in people with a low genetic predisposition to diabetes when compared with people with a high genetic predisposition to diabetes (8,12,13).

In this study, we evaluated the prospective association between accelerometer-measured irregular sleep duration and diabetes risk in a large cohort of U.K. adults. A priori, we hypothesized that people with greater irregular sleep duration would be at higher risk of incident diabetes and that this association would be stronger among participants with a high genetic risk for diabetes or short/long sleep duration.

The UK Biobank (UKB) is an ongoing large community-based prospective cohort study. As of July 2023, the UKB involves 502,364 participants who were 40–69 years old during enrollment in 2006–2010 (14). Participants were recruited from a pool of 9.2 million individuals registered with the National Health Service, representing diverse sociodemographic backgrounds across the U.K. All participants provided e-Consent and attended a baseline visit, which included a touch-screen questionnaire, a computer-assisted interview about sociodemographic information, family history, psychosocial factors, and lifestyle; physical measurements; and blood sample collection (14). Data collection on some baseline measures was repeated for smaller subsamples of participants up to four times. Genotyping and standard biochemical assays were performed at baseline for the entire cohort. Information on health outcomes is routinely collected via linkages to multiple national databases, such as Hospital Episode Statistics data and death registry records.

Between 2013 and 2015, a subsample of 236,519 UKB participants with a valid e-mail address were invited to participate in an accelerometer study (15). With a response rate of 44%, 103,628 participants accepted the invitation and completed the accelerometer measurement. Among 91,451 participants with good quality accelerometer data, we excluded 3,604 participants with prevalent diabetes at the time of accelerometry. We further excluded participants with less than five nights of sleep duration data (n = 2,924), anomalous values (those with sleep duration SD >4 h; n = 34), or missing sleep duration SD values (n = 465), or those lost to medical follow-up before the accelerometer study (n = 3), leaving 84,421 participants for analysis (Supplementary Fig. 1).

Participants wore Axivity AX3 (Open Lab, Newcastle University), a waterproof triaxial accelerometer, on their dominant wrist for 7 days (15). The Axivity AX3 has a stable inter- and intradevice variability when measuring movement, and, in multiaxis shaking tests, it performs slightly better than the GENEActiv accelerometer, a commonly used accelerometer in health research (15,16). Sleep duration variability was defined as the SD of accelerometer-measured sleep duration in hours across main sleep periods. A previous study in UKB (17) derived the mean and SD of accelerometer-measured sleep duration across all days, using an R package, GGIR version 1.5.12 (18). In a validation study among 50 participants, the C statistic for the GGIR package to accurately detect accelerometer-measured sleep period time windows was >0.8 when compared with polysomnography data (19). In the main analysis, we evaluated sleep duration SD across all days both continuously and in categories as ≤30, 31–45, 46–60, 61–90, and ≥91 min.

We applied the same approach as the previous study (17), using GGIR 3.0.9 (20,21), to derive additional accelerometer-based sleep variables for the main sleep period, including sleep onset timing SD across all days and weekdays, mean weekday sleep duration, and weekday sleep duration SD. Additionally, sleep apnea was identified through medical records and self-reports. Insomnia symptoms, chronotype, and daytime sleepiness were self-reported at UKB baseline (Supplementary Table 1).

In the UKB, the first occurrence of diabetes is identified using a standard algorithm, by mapping ICD-10 code E11 (i.e., noninsulin-dependent diabetes mellitus) to participants’ health-related records, including their primary care data, hospital inpatient data, and death register records (22). Additionally, at the UKB assessment center visits, participants self-reported doctor-diagnosed health conditions. A UKB nurse conducted a verbal interview to verify responses, and, when uncertain, a doctor further assessed and classified the self-reported conditions. The algorithm also incorporated verified self-reports of doctor-diagnosed diabetes. This algorithm has been utilized successfully in prior studies to identify incident diabetes cases (23–25). Incident diabetes was defined as diabetes occurring after the accelerometer study end date.

Genotyping was conducted using UKB Axiom (90% of participants) and UK Biobank Lung Exome Variant Evaluation (10% of participants) arrays, directly targeting ∼800,000 genetic markers. An additional 96 million genetic variants were derived using imputation. Rigorous quality control confirmed the high reliability and accuracy of imputed genetic variants (26). In UKB, a Bayesian approach was used to create polygenic risk score (PRS) for diabetes from the meta-analysis of external (non-UKB) genome-wide association studies, and a principal component–based ancestry centering technique was implemented to normalize the diabetes PRS distribution across all ancestries (27,28). We categorized participants according to the PRS as having low (first tertile), intermediate (second tertile), or high (third tertile) genetic risk for diabetes.

We used mean (SD) and percentage to report the distributions for a wide range of covariates across sleep duration SD categories (Supplementary Covariate Assessment and Supplementary Table 1). Cox proportional hazards model was used to estimate the hazard ratio (HR) for the association between sleep duration SD and incident diabetes. Person-time at risk accrued from the accelerometer end-of-wear date up to the onset of diabetes, lost to follow-up, death, or the censoring date (31 May 2022), whichever came first. The proportional hazards assumption was tested by including an interaction term between the exposure and time of follow-up, and no violation of this assumption was detected (P interaction ≥ 0.8608).

We fitted three models. Because comorbidities and adiposity may be a consequence of irregular sleep duration, we evaluated these covariates in separate models (models 2 and 3) to avoid potential overadjustment. In model 1, we included age, sex, and race. In model 2, we further adjusted for Townsend deprivation index, home area type, education, occupation/shift work, family history of diabetes, smoking, physical activity, healthy diet score, coffee or tea intake, alcohol consumption, average noise level, fine particulate matter (PM2.5), and season, along with heart failure, dyslipidemia, hypertension, depression, and osteoarthritis. In model 3, we added BMI and waist circumference. We conducted four sensitivity analyses: 1) We assessed whether the association between sleep duration SD and diabetes risk was independent of other sleep-related factors (sleep apnea, insomnia, mean sleep duration, chronotype, daytime sleepiness, and sleep onset timing SD). 2) To address potential reverse causation, we excluded incident diabetes cases from the first two years of follow-up. 3) To explore whether the association observed for 7-day sleep duration variability was driven by weekday-weekend differences (e.g., as may occur with social jetlag), we repeated the main analysis using sleep duration SD restricted to data collected only for weekdays. 4) As diabetes tended to be underdiagnosed in U.K. during the pandemic (29), we divided the study follow-up into three mutually exclusive periods: the first 2 years of follow-up, from post 2-year follow-up to 11 March 2020 (coronavirus disease 2019 pandemic start date), and after 11 March 2020 (pandemic period). In all analyses, reference group was the most regular sleep duration group (sleep duration SD ≤30 min). We evaluated both the linear trend by modeling sleep duration SD as a continuous variable and potential nonlinearity using restricted cubic splines with three knots at 10th, 50th, and 90th percentiles. Statistical significance of nonlinearity was assessed by a likelihood ratio test comparing the fit of the linear model to the spline model. As our findings suggested a nonlinear relationship, we conducted a post hoc analysis using the nonparametric locally weighted scatterplot smoothing method (30) to identify 62 min as the point of notable curvilinear change. For simplicity, we used the dichotomized sleep duration SD at the 60-min cutoff in further analyses.

In the gene-sleep interaction analysis, we fitted multivariable models as described above, with additional adjustment for the first 10 ancestry principal components to address population stratification. We used an interaction term between PRS and sleep duration SD to evaluate the effect modification on a multiplicative scale. Secondarily, we implemented relative excess risk due to interaction to assess potential effect modification on an additive scale. In parallel, we considered family history of diabetes a second genetic risk indicator for diabetes and repeated the gene-sleep interaction analysis. Participants with a family history of diabetes had a significantly higher PRS (Supplementary Fig. 2). Further, we formulated a composite genetic indicator combining both PRS and family history, following a previous study (31). Participants with both a PRS in the upper tertile and a family history of diabetes were categorized as having a high diabetes genetic risk, while others were categorized as having a low genetic risk. We restricted analyses involving PRS to participants who identified themselves as White.

We performed subgroup analysis, using models 1 and 3, according to accelerometer-measured sleep duration (<7, 7–8, >8 h). We further examined subgroup associations by age at baseline, sex, race, BMI, Townsend deprivation index, occupation/shift work, insomnia symptoms, and chronotype. We assessed the statistical significance of effect modification on a multiplicative scale as described above.

Lastly, we estimated the population attributable risk (PAR) to quantify the proportion of diabetes cases in the population that could be prevented if sleep duration SD was reduced from >60 min to ≤60 min, assuming a causal link between sleep duration SD and diabetes and that the effect of sleep duration SD is completely reversable. We processed and visualized data with Python, and conducted statistical analyses using SAS version 9.4 (Cary, NC).

At the accelerometer study, participants were, on average, aged 62 (SD 8) years, 57% were female, 97% were White, 45% had a college degree, and 50% were employed with nonshift work jobs. Approximately 29% of participants had a sleep duration SD of >60 min (Supplementary Fig. 3). Participants with higher sleep duration SD were younger and more likely to be female, shift workers, or current smokers; report definite “evening” chronotype; have higher Townsend deprivation index (indicating poorer socioeconomic status), higher BMI, and shorter mean sleep duration; and were less likely to be White (Table 1). Family history of diabetes and depression were more prevalent among participants with higher sleep duration SD. Participants with higher sleep duration SD tended to have either short or long mean sleep duration (with more participants toward shorter sleep duration), displaying an asymmetrical U-shaped pattern (P nonlinearity < 0.0001) (Supplementary Fig. 4).

During 622,080 person-years of follow-up (median 7.5 years), 2,058 participants developed diabetes. In model 1, when compared with the participants with ≤30 min of sleep duration SD, the HR (95% CI) for diabetes risk was 1.15 (0.99, 1.33) for 31–45 min, 1.28 (1.10, 1.48) for 46–60 min, 1.54 (1.32, 1.80) for 61–90 min, and 1.59 (1.33, 1.90) for ≥91 min of sleep duration SD (P trend < 0.0001) (Table 2). This association was attenuated in model 2 with adjustment for additional socioeconomic factors, metabolic comorbidities, and depression (P trend = 0.0010), and was further attenuated in model 3 when we additionally controlled for BMI and waist circumference (P trend = 0.0688).

There was evidence for nonlinearity, with participants with >60 min of sleep duration SD consistently having higher risk of diabetes across models (P nonlinearity ≤ 0.0544 in models 1 and 2, and P nonlinearity = 0.2151 in model 3) (Fig. 1). When we dichotomized sleep duration SD, compared with those with ≤60 min of sleep duration SD, participants with >60 min of sleep duration SD had 34% higher diabetes risk (95% CI 1.22, 1.47) in model 1 and 11% higher risk (95% CI 1.01, 1.22) in the fully adjusted model 3. Association patterns were similar when we further accounted for sleep-related covariates (data not shown), excluded diabetes cases diagnosed in the first 2 years of follow-up (Supplementary Table 3), or evaluated the pandemic period separately (Supplementary Table 4). While sleep duration SD was lower across weekdays (mean 1.02; SD 0.56) versus across 7 days (mean 1.10; SD 0.51), we observed somewhat stronger associations when evaluating sleep duration SD across weekdays in relation to diabetes risk (Supplementary Table 5). Lastly, PAR in model 1 was 8.3% (95% CI 5.7, 10.8), suggesting that, with age, sex, and race held constant, 8.3% of diabetes cases in the population could be potentially prevented by reducing sleep duration SD from >60 min to ≤60 min. The corresponding PAR (95% CI) was 1.1% (0.1, 2.1) in model 3.

Sleep duration SD was associated with diabetes risk irrespective of genetic risk for diabetes, although the association was stronger among individuals with lower diabetes PRS (P interaction in model 1: 0.0221) (Table 3). The HR (95% CI) for diabetes risk comparing >60 vs. ≤60 min of sleep duration SD was 1.83 (1.44, 2.31) in participants with low PRS, 1.40 (1.17, 1.67) in intermediate PRS, and 1.21 (1.06, 1.38) in high PRS. Sleep duration SD was not significantly associated with diabetes risk in individuals with intermediate or high PRS after additional covariate adjustment in model 3. However, we did not find evidence of effect modification by family history of diabetes (P interaction ≥ 0.1329) or the composite variable of diabetes PRS and family history (P interaction ≥ 0.2061). When evaluating the joint association between sleep duration SD and PRS, the highest diabetes risk was observed among participants with a high PRS and a sleep duration SD of >60 min compared with a low PRS and sleep duration SD of ≤60 (HR [95% CI] 3.31 [2.75, 3.98]) (Supplementary Fig. 5). No significant interaction on the additive scale between PRS and sleep duration SD on diabetes risk was observed (relative excess risk due to interaction comparing high versus low and intermediate PRS −0.23; 95% CI −0.59, 0.12; P = 0.1997).

The association between sleep duration SD and incident diabetes was stronger among participants with longer hours of mean sleep duration (P interaction ≤ 0.0009) (Table 3). The HR (95% CI) for diabetes risk comparing >60 vs. ≤60 min of sleep duration SD was 1.11 (0.96, 1.28) in participants with mean sleep duration <7 h, 1.41 (1.21, 1.64) in 7–8 h, and 1.65 (1.33, 2.04) in >8 h. We did not find significant differences in subgroup analyses by age, sex, race, Townsend deprivation index tertiles, BMI, insomnia, chronotype, or occupation/shift work (P interactions ≥ 0.0985) (Supplementary Table 6).

We found that irregular sleep duration, measured by 7-day SD in accelerometer-measured sleep duration, was nonlinearly associated with higher risk of developing diabetes. After accounting for comorbidities and adiposity, the association weakened, but remained apparent comparing participants with sleep duration SD of >60 vs. ≤60 min. Contrary to our initial hypothesis, we found that the association was stronger among individuals with low genetic risk for diabetes measured by diabetes PRS, although this gene-sleep interaction was not observed when genetic risk was assessed by family history or the composite measure. Further, the association was more pronounced among individuals with accelerometer-measured long sleep duration versus short sleep duration.

In the largest such study to date, we found that participants with greater sleep duration SD, particularly those with >60 min of sleep duration SD, are at higher risk of developing diabetes than those with lower sleep duration SD. This 60-min cutoff has been used as an empirical cutoff to define irregular sleep duration in other work (33). A 2022 systematic review identified 12 studies that reported inconsistent associations between different sleep variability measures (e.g., sleep regularity index, social jetlag, and sleep duration SD) and diabetes biomarkers, including HbA_1c, insulin, and glucose (10). Another six studies, five of which were cross-sectional (10), evaluated the association between sleep variability and diabetes outcomes (total sample size 10,399). Three found positive associations between sleep irregularity and prevalent diabetes, while the remaining two reported null findings (10). The only prospective study found no association between sleep irregularity and increased diabetes incidence over 5.7 years of follow-up among 2,107 U.S. adults (aged 19–64 years), although a positive cross-sectional association was observed (11). Findings between the current study and the previous prospective study may differ because of variations in race, age, and follow-up periods. Notably, other prospective studies have reported associations between sleep irregularity and cardiometabolic outcomes other than diabetes, such as metabolic syndrome (8,10) and incident cardiovascular diseases (34). Nonetheless, a more recent panel review of existing evidence reached a consensus on the importance of sleep regularity for health, including metabolic health, although it was noted that high-quality, prospective studies were still needed to further characterize the associations (33).

Irregular sleep duration may give rise to diabetes through disrupting circadian rhythms in metabolic regulation and daily patterns of lifestyle behaviors (35). The circadian timing system coordinates various metabolic functions, including insulin secretion and glucose metabolism (2), via alignment with external cues (e.g., light exposure and meal timing). When sleep duration is irregular, these cues are likely to be irregular too, leading to unstable circadian entrainment. Through this mechanism, chronic irregular sleep duration could lead to reduced insulin sensitivity, disrupted glucose metabolism, and development of diabetes and other cardiometabolic outcomes (2,33). Additionally, irregular sleep duration is associated with other metabolically unhealthy sleep traits such as short/long sleep duration, poor sleep quality, and evening chronotype that have been shown to associate with worse glucose metabolism and diabetes risk (36). Adjustment for these sleep-related factors attenuated the association between sleep duration SD and diabetes risk in our analysis, but it did not fully account for it. Further, irregular sleep may interrupt behavioral rhythms and induce unhealthy habits, such as disrupted mealtime, irregular exercise timing (37), and late-night snacking (38), which are emerging risk factors for cardiometabolic disease. Other pathways through which irregular sleep duration may cause cardiometabolic outcomes are inflammation, gut dysbiosis, hypothalamic-pituitary-adrenal axis dysfunction, and autonomic dysfunction (35). Further research is warranted to understand the underlying mechanisms linking irregular sleep duration with diabetes risk (33).

Irregular sleep duration was associated with increased diabetes risk regardless of the diabetes PRS, suggesting that promoting regular sleep duration may benefit individuals with low, intermediate, or high genetic susceptibility for diabetes. In contrast to our hypothesis, we found stronger associations in individuals with low PRS, while individuals with intermediate or high PRS showed no significant associations after additionally adjusting for comorbidities and lifestyle factors. It is possible that the impact of sleep regularity on diabetes may be masked in the presence of a high genetic risk. However, it should be noted that such gene-sleep interaction effects were not consistently observed across different scales and gene-related variables. While the multiplicative interaction was observed for diabetes PRS, it was not for family history and the composite measure. Additionally, for diabetes PRS, we did not find evidence of an interaction on the additive scale. Similarly, a 2015 meta-analysis of 15 cohort studies did not identify a robust gene-sleep duration interaction on cardiometabolic traits (39). However, we were unable to identify studies on the interactions between genetic susceptibility and irregular sleep duration. We further found stronger sleep duration SD and diabetes association among participants with >8 h of average sleep duration. Longer sleep duration might reduce daylight exposure, which could, in turn, give rise to circadian disruption. To further evaluate the role of genetic risk, future studies could apply causal inference methodology considering risk factor variations across strata of genetic risk.

Our study has several key strengths. First, we utilized a large, population-based sample, which reduces sampling error. Second, we utilized sleep data that were objectively collected in participants’ usual environment, and used a validated method for identifying diabetes cases through medical records and subsequent clinical review. Third, our use of a prospective study design, in contrast to a cross-sectional approach, and careful adjustment for many important covariates provides enhanced causal insights into the association. Lastly, we employed a metric for sleep variability that is applicable in broader contexts and might allow for simpler interpretation compared with other measures. While multiple metrics exist and are often highly correlated, there is currently no consensus on a universal metric. However, compared with other metrics like sleep regularity index, sleep duration SD has more practical interpretations that may be converted to quantitative public health recommendations for improving sleep regularity (33,40).

Our study has some limitations. Most importantly, the time lag between covariate assessment and sleep duration SD measurements for some participants was large, ranging from 0 to 9 years (median 5 years). While this may not affect stable covariates like demographics or family history, it could bias lifestyle behaviors that can vary over time. To reduce potential residual confounding due to these time intervals, we updated the confounder information for participants who completed follow-up assessment visits after baseline (n = 9,542) by using the one conducted closest to the accelerometer study date. Further, sensitivity analyses suggest high stability for BMI values measured, on average, 4.2 years apart (r = 0.93). Another important limitation relates to the assessment of sleep duration SD based on 7-day sleep duration, which may not capture long-term habitual sleep patterns for some participants. Moreover, response rate in parental UKB cohort was ∼5.5% (14), and participants in the accelerometer study were selected based on a constrained random sampling approach, raising the potential of selection bias. However, risk factor associations found using UKB data were comparable to representative studies with conventional response rates (41). Lastly, our findings have limited generalizability, because of the relatively healthy, predominantly White study sample. Further research is essential to determine the applicability of our results to other populations.

In conclusion, our prospective study indicates that middle-aged to older adults with inconsistent sleep duration over the week have a heightened risk of developing diabetes relative to their counterparts with more consistent sleep patterns. This association persisted even after rigorous covariate adjustment, including obesity, comorbidities, family history of diabetes, and various lifestyle factors. Notably, individuals exhibiting a 1-week sleep duration variability exceeding 60 min were found to have consistently elevated diabetes risk compared with those with a variability of less than 60 min, suggesting a 1-h cutoff to define metabolically unhealthy irregular sleep duration. Further intervention studies and mechanistic research are needed to fully understand the biological and behavioral pathways linking irregular sleep duration to increased diabetes risk.

Acknowledgments. The authors express their gratitude to the participants and staff of the UKB study for their valuable contributions to research. During the course of preparing this work, the author(s) used OpenAI DALL-E for the purpose of making the image of “a person sleeping with an accelerometer” in the Graphical Abstract. Following the use of this tool/service, the author(s) formally reviewed the content for its accuracy and edited it as necessary.

Funding. This study was supported by the National Institutes of Health (grant number R01HL155395) and the UKB project 85501. S.K. was supported by the American Heart Association Postdoctoral Fellowship (https://doi.org/10.58275/AHA.24POST1188091.pc.gr.190780).

Duality of Interest. A.P. has received research funding from Versalux and Delos, and he is a director and founder of Circadian Health Innovations PTY LTD. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. T.H. supervised the work. S.K. and T.H. contributed to conception, study design, data acquisition, and analysis. S.K. wrote the first draft of the manuscript, and all authors edited, reviewed, participated in data interpretation, and approved the final version of the manuscript. S.K. and T.H. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at SLEEP 2023, Indianapolis, IN, 3–7 June 2023.

Handling Editors. The journal editors responsible for overseeing the review of the manuscript were Steven E. Kahn and Justin B. Echouffo-Tcheugui.