Associations between sleep disturbances, diabetes and mortality in the UK Biobank cohort: A prospective population‐based study

Abstract Non‐communicable diseases, including diabetes, are partly responsible for the deceleration of improvements of life expectancy in many countries. Diabetes is also associated with sleep disturbances. Our aim was to determine whether sleep disturbances, particularly in people with diabetes, were associated with increased mortality risk. Data from the UK Biobank were analysed (n = 487,728, mean follow‐up time = 8.9 years). The primary exposure was sleep disturbances, assessed through the question: Do you have trouble falling asleep at night or do you wake up in the middle of the night? The primary outcome was mortality. We also dichotomized sleep disturbances into “never/sometimes” versus “usually” (frequently), and combined with the presence/absence of diabetes: 24.2% of participants reported “never/rarely” experiencing sleep disturbances, 47.8% “sometimes” and 28.0% “usually”. In age‐ and sex‐adjusted models, frequent sleep disturbances were associated with an increased risk of all‐cause mortality (hazard ratio [HR], 1.31; 95% confidence interval [CI], 1.26–1.37), which remained significant in the fully adjusted model (HR 1.13, 95% CI 1.09–1.18). The presence of both diabetes and frequent sleep disturbances was associated with greater risk of all‐cause mortality than either condition alone. In the fully adjusted model, the hazard ratio for all‐cause mortality was 1.11 (95% CI, 1.07–1.15) for frequent sleep disturbances alone, 1.67 (95% CI, 1.57–1.76) for diabetes alone and 1.87 for both (95% CI, 1.75–2.01). Frequent sleep disturbances (experienced by more than one quarter of the sample) were associated with increased risk of all‐cause mortality. Mortality risk was highest in those with both diabetes and frequent sleep disturbances. Complaints of difficulty falling or staying asleep merit attention by physicians.


Increases in life expectancy have slowed or even ceased in the United
States, the United Kingdom and comparable countries (Murphy, Xu, Kochanek, & Arias, 2017;Office for National Statistics, 2018). In the majority of countries, people have a high risk of premature mortality due to non-communicable diseases (NCDs), including diabetes, compared with other conditions (NCD Countdown 2030Collaborators, 2018). The mortality rates due to these NCDs are expected to increase by approximately 54% between 2016 and 2040, and deaths due specifically to type 2 diabetes are estimated to more than double worldwide (Foreman, Marquez, & Dolgert, 2018). Indeed, diabetes increases the risk of all-cause and cardiovascular mortality (Haffner, Lehto, Ronnemaa, Pyorala, & Laakso, 1998). An international group led by the United Nations set a goal of reducing the rates of premature mortality due to NCD by one third by the year 2030, but only 16% of countries are on target for men and 19% of countries are on target for women (NCD Countdown 2030Collaborators, 2018.
Given that NCDs are the leading causes of death and the rates of premature mortality rates are not declining as targeted, it is important to gain a greater understanding of the underlying causes of NCDassociated mortality. Here, we have used data from the UK Biobank to examine the effect of sleep disturbances and their interactions with diabetes, a major NCD, on morbidity and mortality.
The aims of the analyses presented here were to determine whether sleep disturbances were associated with increased risk of all-cause and cardiovascular disease (CVD) mortality in a large study of adults in the UK and to determine whether having both frequent sleep disturbances and diabetes was more strongly associated with mortality than either condition on its own. We hypothesized that frequent sleep disturbances would be associated with increased mortality risk, particularly in those with diabetes.

| ME THODS
We used data from the UK Biobank, which is a large, prospective, population-based cohort study designed to investigate risk factors for major diseases of middle and older age (Sudlow et al., 2015). It enrolled 502,642 people aged 37-73 years (53% women) from across the UK between March 2006 and October 2010. The study invited every individual within this age range who was registered with the National Health Service and living up to about 25 miles from an assessment centre to participate (Allen et al., 2012). Identical assessment procedures were used across field sites. For the analyses presented here, we have death records up to 14 February 2018, which resulted in a mean follow-up time of 8.9 years (range, 4 days to 11.9 years) among participants and that 98% of deaths occurred after 6 months of follow-up.
The UK Biobank protocol is available online (Palmer, 2007).
The primary exposure variable was the presence of sleep disturbances, as assessed through a single question that asked: Do you have trouble falling asleep at night or do you wake up in the middle of the night? There were three response options: "never/rarely", "sometimes" or "usually". Only 0.1% of participants did not answer this question.
The primary outcome was mortality, including all-cause mortality and mortality due to CVD. Mortality information was obtained from the National Health Service for England and Wales and the NHS Central Register in Scotland. All details from the death certificates were provided to UK Biobank personnel, who coded primary cause of death according to ICD10. We classified the ICD10 codes I00-I99 as CVD-related mortality. If no death was recorded for a participant, they were assumed to still be living.
We examined the interaction between the presence of sleep disturbances and the presence of diabetes. Participants were classified as having diabetes if they reported a previous diagnosis of diabetes or they reported taking insulin. We dichotomized sleep disturbances into "never"/"sometimes" (infrequently) versus "usually" (frequently) and combined this with presence/absence of diabetes to create four groups. We also created six groups based on three levels of frequency of sleep disturbance and presence of diabetes.
Covariates included age, sex, ethnicity, socioeconomic deprivation, smoking, sleep duration, body mass index (BMI) and comorbidities. Age was calculated based on date of birth and date of examination. Sex was acquired from the central registry and updated by the participant, if needed. Ethnicity was self-identified and the majority of the sample (94%) described themselves as "white". Therefore, we dichotomized ethnicity into "white" and "non-white".
The measure of socioeconomic deprivation was based on the Townsend deprivation index (Townsend et al., 1988), which summarizes deprivation in a postcode area based on the rates of unemployment, absence of ownership of a car and home, and household overcrowding. Smoking status was obtained by self-report with the following categories: "never", "previous smoker", "current smoker" and "prefer not to answer". Sleep duration was based on the question: About how many hours sleep do you get in every 24 h (please include naps)? Responses were provided as integers. Standing height was measured using a Seca 240-cm height measure while participants stood barefoot with posture verified by trained staff. Weight was measured using a Tanita BC418MA body composition analyser and BMI (kg/m 2 ) was calculated. Comorbidities were recorded based on self-report during an interview by a trained nurse and all comorbidity variables are dichotomous (present/absent). We used the codes provided by the UK Biobank to classify them into the following 11 comorbidity variables for analyses (see Table S1 for complete list of codes): CVD, diabetes, other endocrine disorders, neurological disorders, renal disorders, respiratory disorders, musculoskeletal disorders, gastrointestinal/abdominal disorders, depression, other psychological disorders and cancer. A participant only needed a report of one code to be classified as having that comorbidity. Each comorbidity was treated as a separate variable in the analyses.
Insomnia symptom data were missing for 1504 participants, 9705 participants were missing BMI, 3000 were missing sleep duration, 603 were missing the Townsend index and three participants had a negative follow-up period, resulting in a final sample size of 487,728 participants. Means and standard deviations (SD) or proportions (%) were calculated to describe the sample. Cox proportional hazards models were estimated to determine associations with allcause and CVD mortality. There were two models, one adjusting for age and sex only and one adjusting for all covariates. Further, we examined sleep disturbances alone ("never/rarely" was referent) as well as the four or six groups based on frequency of sleep disturbances and diabetes (absence of both was referent). In addition, to determine whether the two groups with diabetes differed, we repeated the analysis with the four groups using the diabetes-alone group as the referent. Finally, we tested an interaction term between diabetes and sleep disturbances to determine whether the association between sleep disturbances and mortality risk varied between those with and without diabetes. All statistical analyses were performed using Stata, v14 (Stata corp).

TA B L E 1 Description of full sample and by insomnia symptom frequency
compared to those who had neither condition (Figure 1). In the fully adjusted model, the hazard ratio for all-cause mortality was

| D ISCUSS I ON
In this large, UK-based population study, we observed significant associations between frequent sleep disturbances and risk of allcause mortality. Frequent sleep disturbances were experienced by more than one quarter of the sample, and thus are highly prevalent in the UK, which is consistent with other observational studies (Ohayon, 2002). In addition, individuals with diabetes who also experienced frequent sleep disturbances had a greater risk of mortality than those with diabetes who did not report frequent sleep disturbances. Further, the association between frequent sleep disturbances and mortality risk did not differ between those who did and did not have diabetes.
Our findings from this UK cohort are consistent with studies from other countries. A study from Norway assessed self-reported insomnia in adults aged 40-45 years at baseline and followed them for 13-15 years. They reported that insomnia at baseline was a significant predictor of all-cause mortality (HR, 3.34; 95% CI, 1.67-6.69) (Sivertsen et al., 2014). A large study of middleaged and older men in the US also reported that difficulty initiating sleep (HR, 1.55; 95% CI, 1.19-2.04) and non-restorative sleep (HR, 1.32; 95% CI, 1.02-1.72) were associated with increased risk of all-cause mortality (Li et al., 2014). A Chinese study observed that adults who reported sleep disturbances nearly every day had an increased risk of mortality over approximately 16 years (Chien et al., 2010). A community-based prospective study in the USA with a 20-year follow-up found that persistent insomnia (reporting symptoms at two assessments) was associated with increased mortality risk; however, reporting sleep disturbances at only one assessment was not (Parthasarathy, Vasquez, & Halonen, 2015).
The discrepant findings could be due to differences in insomnia assessment, characteristics of the population studied or covariate adjustment, including comorbidities.
Sleep disturbances have been associated with CVD in prior research. For example, one prospective population-based study from Norway followed participants for approximately 11 years and found that risk of acute myocardial infarction was significantly higher for individuals who had difficulty falling asleep almost every night (HR, 1.45; 95% CI, 1.18-1.80) and for individuals who had difficulties maintaining sleep almost every night (HR, 1.30; 95% CI, 1.01-1.68) compared to those who never have these sleep difficulties (Laugsand, Vatten, Platou, & Janszky, 2011). The same study also observed a significant increased risk of heart failure among those who had difficulty falling asleep almost every night (HR, 1.32; 95% CI, 1.01-1.72) compared to those who never have these sleep difficulties (Laugsand, Strand, Platou, Vatten, & Janszky, 2014). A large population-based study in Taiwan also observed a significant increased risk of acute myocardial infarction, as well as stroke, among people with diagnosed insomnia (Hsu et al., 2015). In our study, however, we did not observe a significant association between frequent sleep disturbances and CVD mortality during the 8.9-year follow-up period. This may be because mortality from CVD is not impacted by sleep disturbances (at least as defined by the single question used here or limited power due to lower number of CVD deaths).
To our knowledge, this is the first study to examine the effect of the combination of insomnia and diabetes on mortality risk. Diabetes has been previously associated with increased risk of CVD and mortality (Haffner et al., 1998;Stamler, Vaccaro, Neaton, & Wentworth, 1993), and diabetes has been associated with impaired sleep. Several studies have found a strong association between obstructive sleep apnea (OSA) and type 2 diabetes (Huang et al., 2018;Subramanian et al., 2019), and OSA impairs sleep quality. Further, some observational studies have found that among people with type 2 diabetes, worse sleep quality is associated with higher haemoglobin A1c, suggesting poorer glycaemic control (Knutson et al., 2006(Knutson et al., , 2011. A meta-analysis of nine studies among adults with type 2 diabetes also found that poor sleep quality was associated with higher haemoglobin A1c (Lee, Ng, & Chin, 2017), and insomnia has been identified as a risk factor for type 2 diabetes both in observational (Vgontzas et al., 2009) and Mendelian randomization studies (Yuan & Larsson, 2020). Experimental studies that impaired sleep quality did observe impairments in glucose regulation in healthy volunteers (Stamatakis & Punjabi, 2010;Tasali, Leproult, Ehrmann, & Van Cauter, 2008). If chronic poor sleep quality due to a sleep disorder can impair glucose control in people with diabetes, then this could be a mechanism leading to the increased risk of mortality in people with diabetes and frequent sleep disturbances.
The strengths of this study include the large sample size and the prospective monitoring of mortality. The UK Biobank study aimed to assemble a general population sample; however, this cohort does appear to be slightly healthier on average than the general UK population (Fry, Littlejohns, & Sudlow, 2017), which may limit generalizability somewhat. We also do not have access to measures of hypnotic or alcohol use in our dataset and these could be important confounders or mediators of the association between sleep disturbances and mortality. Another limitation is that the mean follow-up time is only 8.9 years and a longer period would result in a higher number of mortalities, which could increase power for the CVD mortality analyses. Finally, sleep disturbances are based on a single self-reported question, which did not assess daytime consequences, and are not equivalent to a clinically diagnosed insomnia disorder. However, this same question in the same sample has recently been used successfully for genomewide association studies (Jansen, Watanabe, & Stringer, 2019;Lane et al., 2016), with the most significant hits being replicated both in a sub-stratification of the UK Biobank study based on accelerometry and in a separate insomnia cohort (Lane et al., 2016). In addition, people who report sleep disturbances are likely to be a heterogeneous clinicians to consider, particularly for diabetes patients. We found