Risk of Incident Obstructive Sleep Apnea Among Patients With Type 2 Diabetes

Obstructive sleep apnea (OSA) is a common disorder characterized by repeated complete or partial upper airway obstructions during sleep, leading to recurrent oxygen desaturations, cyclical adverse changes in heart rate, blood pressure (BP), and sympathetic activity, and disruption to sleep architecture (1). As a result, OSA has been linked to multiple adverse outcomes, including road traffic accidents (2), reduced workplace productivity (3), cardiovascular disease (CVD) (4,5), hypertension (5,6), insulin resistance (7), increased mortality (8), and type 2 diabetes (9).

Multiple epidemiological studies have indicated that OSA may be a risk factor for the development of type 2 diabetes independent of obesity and other confounders (9), possibly via the effect of recurrent hypoxemia, inflammation, sympathetic activation, and activation of the hypothalamic-adrenal axis on insulin resistance and β-cell function (10). In addition, multiple cross-sectional studies have shown a high prevalence of undiagnosed OSA in patients with type 2 diabetes (24–86%) (11,12). Furthermore, cross-sectional and cohort studies have shown that OSA is associated with worse glycemic control in patients with type 2 diabetes (11–13) and with diabetes-related vascular complications (14–17). Hence, understanding the complex relationships between OSA and type 2 diabetes is important, because OSA might be considered a modifiable risk factor for type 2 diabetes development and adverse outcomes in these patients (18).

However, the potential role for type 2 diabetes as a risk factor for OSA has been less well examined. The relationship between OSA and type 2 diabetes is plausibly bidirectional. Type 2 diabetes might lead to the development of OSA as a result of the effect of insulin resistance and autonomic dysfunction on upper airways stability (19–21) as well as the increase in weight associated with some type 2 diabetes treatments. Hence, there is a need to examine the effect of type 2 diabetes on incident OSA. Identifying predictors of OSA in patients with type 2 diabetes is also important to aid screening strategies.

The primary aim of our study was to evaluate the role of type 2 diabetes as a risk factor for incident OSA by using a matched controlled population-based retrospective cohort study. The secondary aim was to identify predictors of incident OSA in patients with type 2 diabetes.

An age-, sex-, and BMI-matched controlled retrospective cohort study was conducted from 1 January 2005 to 31 December 2017.

Data were extracted from The Health Improvement Network (THIN), an electronic primary care records database that contains anonymized medical records of more than 15 million patients from 787 practices in the U.K. The database is generalizable to the U.K. population. It consists of coded information on patient demographics, symptoms and diagnoses, drug prescriptions, consultations, and diagnostic tests and their results. THIN is particularly suitable for analyzing long-term health outcomes because general practitioners routinely collect and coordinate the patient’s data (22). THIN has been extensively used previously to study cardiometabolic outcomes (23,24) and to study type 2 diabetes and OSA (25).

The THIN data collection scheme and research performed using THIN data were approved by the National Health Service South-East Multicentre Research Ethics Committee in 2003. Under the terms of the approval, studies must undergo independent scientific review. The Scientific Review Committee approved the use of the use of THIN data for this study in July 2018 (SRC reference 18THIN062).

To ensure high-quality data, general practices were eligible for inclusion in the study from the latest of 12 months after reporting acceptable mortality rates, 12 months after starting to use electronic medical records, and the study start date (1 January 2005). Adult patients aged 16 years and older registered for at least 12 months with any of the eligible general practices before study entry formed the source population.

The exposed cohort consisted of adult patients with type 2 diabetes. Type 2 diabetes diagnosis was ascertained by the presence of any type 2 diabetes clinical (Read) code (Supplementary Table 1A) in the patient’s medical record and the absence of any record of type 1 diabetes. The Read code list used to define exposure has previously been used to study type 2 diabetes (23). The primary analysis included all patients (prevalent and incident) with a type 2 diabetes diagnosis. A sensitivity analysis, including only patients with incident type 2 diabetes (newly diagnosed during the study period), was performed to explore any effect of survival bias.

For every exposed patient, up to four control subjects were randomly selected from an age-, sex-, and BMI-matched pool of eligible patients without a record of type 2 diabetes at any time before or during the study period. Age and BMI were matched to within 1 year and 2 kg/m² respectively.

A 15-month latency period was used for all patients. For patients with incident type 2 diabetes, the index date was 15 months after the date of diagnosis; for patients with prevalent type 2 diabetes, the index date was 15 months after the date the patient became eligible for inclusion. The 15-month interval was introduced to 1) ensure that all predictors determining the risk of OSA in patients with diabetes were recorded at baseline, because the Quality and Outcomes Framework (QOF) ensures these are captured within a 15-month period; and 2) limit the possibility of silent OSA preceding type 2 diabetes being misclassified as incident OSA. The unexposed patients were assigned the same index date as their corresponding exposed patient to avoid immortal time bias (26). Patients with type 2 diabetes and control subjects were monitored from the index date until the earliest of the following end points: outcome (OSA) date, death date, date patient left the practice, date the practice ceased contributing to the database, and study end date (31 December 2017).

OSA was identified by a record of any relevant clinical code (Supplementary Table 1B).

Poisson regression was used to calculate crude incidence rate ratios (IRRs) and adjusted incidence rate ratios (aIRRs), together with their corresponding 95% CIs, 1) comparing the incidence of OSA in patients with and without type 2 diabetes and 2) in an analysis restricted to patients with type 2 diabetes to explore possible risk factors that may predict the incidence of OSA. These analyses excluded patients with a record of the outcome at baseline.

Further exploratory analysis comparing the incidence of OSA between patients with and without type 2 diabetes was performed in subgroups of patients stratified by age, sex, BMI, and the presence or absence of comorbid conditions, including composite CVDs composed of heart failure, ischemic heart disease, stroke/transient ischemic attack (TIA), atrial fibrillation, and hypertension, and composite mental health conditions (anxiety and depression).

To evaluate any possible effect of surveillance bias and unobserved confounders on the outcome (OSA), in the same cohort we estimated the aIRR of chronic obstructive pulmonary disease (COPD), a negative control outcome that has symptoms that overlap with those of diabetes (sleep disturbance and fatigue).

A further sensitivity analysis was performed restricting outcomes to those specifically recorded as obstructive-type sleep apnea (Read code Fy03.11 or H5B0.00) to ensure the exclusion of central sleep apnea outcomes.

Covariates for the study were selected based on biological plausibility and previous literature. All regression models were adjusted for age, sex, BMI, Townsend deprivation quintile, smoking status, and ethnicity. BMI recorded closest to the index date was categorized as <25 kg/m², 25–30 kg/m² (overweight), and ≥30 kg/m² (obese). Implausible BMI values <14 and >75 were considered missing. Social deprivation was categorized as quintiles based on the Townsend score (27). Smoking status was categorized as nonsmoker, previous smoker, and smoker. Missing values for BMI, Townsend deprivation quintile, and smoking status were treated as a separate missing category.

In the model identifying risk factors for OSA among patients with type 2 diabetes, further diabetes-related covariates recorded before the index date were taken into consideration. These included HbA_1c category, estimated glomerular filtration rate (eGFR) category, record of hypoglycemic attack, diagnosis of foot disease, retinopathy, CVD (heart failure, ischemic heart disease, stroke/TIA, atrial fibrillation, and hypertension), mental health conditions, and prescription of lipid-lowering drugs and insulin within 60 days before the index date. Implausible HbA_1c values >200 mmol/mol and eGFR >200 mL/min/1.73 m² were considered missing. Lipid-lowering drugs were used as a proxy measure for hypercholesterolemia and insulin as an indicator of diabetes severity.

eGFR was calculated using the Chronic Kidney Disease Epidemiology Collaboration equation from the serum creatinine values and ethnicity data (where available). HbA_1c was categorized as ≤6.5% (47.50 mmol/mol), 6.5–7.5% (47.51–58.50 mmol/mol), 7.5–8.5% (58.51–69.40 mmol/mol), and >8.5% (69.40 mmol/mol); calculated eGFR was categorized as ≥90 mL/min/1.73 m² (stage 1 chronic kidney disease), 60–89 mL/min/1.73 m² (stage 2), 30–59 mL/min/1.73 m² (stage 3), and <30 mL/min/1.73 m² (stage 4 and 5). Retinopathy was considered as a composite of sight-threatening retinopathy graded as R2 (preproliferative), R3 (proliferative), or M1 (maculopathy), vision loss, and use of medical procedures such as laser and vitreous injections. Diabetes-related foot disease was a composite of lower-limb amputation, gangrene, foot ulcer, Charcot foot, peripheral vascular disease, and peripheral neuropathy. The covariates were identified by clinical codes indicating the condition or treatments performed specific to the condition. Most covariates are included in disease registers that general practices are expected to maintain according to the QOF (28).

Analyses were performed in Stata IC 14 software. Two-sided P values were obtained, and a P value <0.05 was considered as statistically significant.

The results of this study are reported in line with RECORD (REporting of studies Conducted using Observational Routinely-collected health Data) guidelines (Supplementary Table 2).

We identified 360,250 eligible patients with type 2 diabetes. These patients were matched for age, sex, and BMI to 1,296,489 patients without type 2 diabetes (unexposed/control cohort). Baseline characteristics are reported in Table 1. The matching parameters age and sex were similar between the exposed and unexposed groups (mean [SD] age, 64.9 [13.3] vs. 64.6 [13.6] years; male sex, 55.5% vs. 54.2%). Patients in the exposed cohort had a slightly higher mean BMI than control patients (31.0 [6.5] vs. 29.8 [5.8] kg/m²), but the difference was within the matching range (±2 kg/m²). Compared with control patients, patients with diabetes were more deprived (13.7% vs. 9.9% were in the most deprived Townsend quintile) and more likely to be of south Asian ethnicity (3.8% vs. 0.9%). Patients with diabetes also had higher levels of CVDs, including heart failure (4.8% vs. 2.5%), ischemic heart disease (19.1% vs. 11.4%), and stroke/TIA (8.8% vs. 5.9%), and greater use of lipid-lowering drugs (63.7% vs. 23.6%). Prevalent OSA at baseline (recorded up to 15 months after the index date) was higher among exposed compared with unexposed patients (1.8% vs. 0.9%); these patients were excluded in subsequent analyses.

OSA developed in 3,110 patients (0.88%) with diabetes and in 5,968 control patients (0.46%) during the follow-up period (Table 2). The crude incidence rate of OSA among patients with and without diabetes was 1.76 and 1.00 per 1,000 person-years: IRR 1.76 (95% CI 1.69–1.84; P < 0.001). After adjustment for potential confounders, including age, sex, BMI, Townsend deprivation quintile, smoking status, and ethnicity, the association remained statistically significant: aIRR 1.48 (95% CI 1.42–1.55; P < 0.001). Further adjustment for composite CVD at baseline slightly attenuated the effect: aIRR 1.36 (95% CI 1.30–1.42; P < 0.001). The association remained similar in a further sensitivity analysis including only patients with incident type 2 diabetes and their corresponding control patients (Supplementary Table 3A): aIRR 1.41 (95% CI 1.32–1.51; P < 0.001) (Table 2).

With the exception of patients aged 16–29 years, in all of the age- and BMI-stratified cohorts there was a significant increase in OSA incidence in patients with type 2 diabetes compared with patients without (Fig. 1). The risk of OSA was increased in both sexes, but the association between diabetes and OSA was stronger in women (aIRR 1.76 [95% CI 1.62–1.91]) than in men (1.39 [1.32–1.46]). Irrespective of the presence or absence of comorbid conditions, there was a statistically significant increase in OSA incidence in patients with type 2 diabetes.

In the analysis restricting outcomes to those explicitly coded as obstructive-type sleep apnea, there was an increase in the observed effect size: aIRR 1.62 (95% CI 1.52–1.73; P < 0.001) (Supplementary Table 3B).

The analysis considering COPD as an outcome found no significant increase in the incidence of COPD in patients with type 2 diabetes compared with patients without type 2 diabetes after adjusting for age, sex, BMI, Townsend deprivation quintile, smoking status, and ethnicity: aIRR 1.03 (95% CI 0.99–1.08; P = 0.112).

Among the 20 risk factors considered, male sex (aIRR 2.27 [95% CI 2.09–2.46]; P < 0.001), being overweight (2.02 [1.54–2.64]; P < 0.001) or obese (8.29 [6.42–10.69]; P < 0.001), being a previous smoker (1.13 [1.04–1.22]; P = 0.004), diabetes-related foot disease (1.23 [1.06–1.42]; P = 0.005), heart failure (1.41 [1.18–1.70]; P < 0.001), ischemic heart disease (1.22 [1.11–1.34]; P < 0.001), atrial fibrillation (1.23 [1.04–1.46]; P = 0.015), hypertension (1.32 [1.23–1.43]; P < 0.001), depression (1.75 [1.61–1.91]; P < 0.001), and insulin prescription (1.58 [1.42–1.75]; P < 0.001) were significantly predictive of incident OSA in patients with type 2 diabetes (Table 3).

In a sensitivity analysis including only patients with incident/newly diagnosed type 2 diabetes, the results remained similar except for four of the risk factors considered: hypoglycemic event became statistically significant as a predictor of OSA (aIRR 2.06 [95% CI 1.26–3.39]; P = 0.004), whereas being a previous smoker, atrial fibrillation, and prescription of insulin became nonsignificant as predictors in the model (Supplementary Table 3C).

This study showed that patients with type 2 diabetes have an almost 50% increase in risk of developing OSA compared with patients without type 2 diabetes, independent of potential confounders and traditional OSA risk factors. In addition, our study identified predictors of incident OSA in patients with type 2 diabetes, including male sex, obesity, CVDs, diabetes-related foot disease, insulin use, and depression. This is the first study to identify predictors of incident OSA in patients with type 2 diabetes and to examine the links between type 2 diabetes and incident OSA in a European population.

Previous literature has shown that OSA is an independent risk factor for the development of type 2 diabetes. Our results suggest that this relationship is bidirectional, because patients with type 2 diabetes were also at increased risk of developing OSA, despite excluding patients in whom OSA was diagnosed up to 15 months after the type 2 diabetes diagnosis. Although many studies have examined the effect of OSA on type 2 diabetes, there is little evidence regarding the effect of type 2 diabetes on OSA. One longitudinal cohort study of 1,780 men and 1,785 women, the Data from an Epidemiologic Study on the Insulin Resistance Syndrome (D.E.S.I.R.) study, found that fasting insulin (odds ratio 1.31; 95% CI 1.13–1.51) and HOMA of insulin resistance (odds ratio 1.24; 95% CI 1.09–1.4) were predictors of incident “witnessed apnea” over a 6-year period independent of obesity (19). This result suggests that dysglycemia and insulin resistance might lead to the development of OSA, but there was no formal assessment of OSA in that study because the diagnosis was based on self-reported witnessed apneas during sleep.

A more recent analysis of the combined population of 146,519 participants from the Nurses’ Health Study (NHS) (2002–2012), Nurses’ Health Study II (NHSII) (1995–2013), and Health Professionals Follow-up Study (HPFS) (1996–2012) by Huang et al. (29) showed that patients with type 2 diabetes were at increased risk of developing OSA compared with patients without diabetes (hazard ratio [HR] 1.53; 95% CI 1.32–1.77) which was attenuated after adjustment for obesity (HR 1.08; 95% CI 1.00–1.16). Consistent with our findings, Huang et al. also found that insulin-treated patients with type 2 diabetes were at increased risk of OSA compared with patients without diabetes after adjustment for obesity (HR 1.43; 95% CI 1.11–1.83), particularly among women (HR 1.60; 95% CI 1.34–1.89).

Our study differs from the Huang et al. (29) study in multiple aspects. The Huang et al. study included patients who were free of CVD at baseline, whereas CVD was common in our study population, suggesting that we have a population with more advanced disease. The diagnosis of OSA in the Huang et al. study was based on self-reporting, whereas the OSA diagnosis in our study was based on clinical codes indicating physician diagnosis. The design of the two studies also differed, because our study matched for major OSA risk factors between patients with and without type 2 diabetes. Finally, the relationship between type 2 diabetes and incident OSA became nonsignificant when adjusting for obesity in the Huang et al. study, whereas our study showed that type 2 diabetes predicted incident OSA independent of obesity and despite adjustment for a wider range of potential confounders than those considered in the study by Huang et al. Despite these differences, our findings, together with those of the above-mentioned studies, strongly suggest that the relationship between type 2 diabetes and OSA is bidirectional.

The lack of a linear relationship between age and OSA incidence in patients with type 2 diabetes in this study is interesting and possibly explained by the fact that the average age of our study patients with type 2 diabetes was ∼65 years old, and several previous studies have shown that the age-related increase in prevalence occurred before age 65 years (30), a finding that is reflected in our analysis.

In age-, sex-, and BMI-stratified subgroup analyses, increased incidence of OSA was observed in patients with type 2 diabetes in all of the subgroups except among patients aged 16–29 years. There was a greater effect size in women compared with men, which concurs with previous evidence suggesting an increased susceptibility of women with type 2 diabetes to adverse health outcomes (31). The observed effect sizes were slightly higher in patients with comorbid conditions, but this was not statistically significant. In a sensitivity analysis including baseline cardiovascular conditions as a covariate, there was a slight reduction in the effect size, indicating that CVD may be a potential effect modifier in the association between type 2 diabetes and incident OSA.

To explore any effect of surveillance bias, we performed an exploratory analysis considering COPD as the outcome instead of OSA. COPD was used because increased screening rates and surveillance, and therefore detection, might occur in patients with type 2 diabetes because the two conditions have common symptoms, including sleep disturbance and fatigue. However, there was no statistically significant increase in incident COPD in patients with type 2 diabetes, suggesting that the observed increase in incident OSA reflects a true difference.

Although the effect of OSA on incident type 2 diabetes is probably related to increased oxidative stress, inflammation, sympathetic activation, and hypothalamic-pituitary-adrenal activation (32), type 2 diabetes might lead to the development of OSA via multiple mechanisms, including weight gain mediated by medications, such as sulfonylureas, glitazones, and insulin, or by hypoglycemia-decreased physical activity, by comorbidities such as diabetic neuropathy, or by changes related to lung volumes (10). Our study supports some of these plausible mechanisms by showing obesity, diabetes-related foot disease (which encompasses neuropathy), and insulin treatment were predictors of incident OSA in patients with type 2 diabetes. However, the relationship between type 2 diabetes and OSA persisted after adjustment for a wide range of potential confounders. Interestingly, depression and CVD were predictors of incident OSA in our study, which aligns with previous studies showing high prevalence of OSA in these conditions (33,34). However, some traditional OSA risk factors, such as age, ethnicity, and smoking, were not identified as predictors of OSA in our study. This could potentially be due to the multiple risk factors considered in the model, where weak effects were obscured, or because our study population included only patients with type 2 diabetes rather than the general population.

Because of the high prevalence of OSA in patients with type 2 diabetes, the International Diabetes Federation (IDF) issued a statement suggesting that all patients with type 2 diabetes should be screened for OSA (35). However, owing to the large number of patients with type 2 diabetes and the lack of easily available screening methods beyond questionnaires, this IDF statement has not been widely followed (36). Our study supports the IDF statement by identifying high-risk groups within patients with type 2 diabetes who are at particularly high risk of developing OSA, including men and patients with CVD, depression, diabetes-related foot disease, patients experiencing hypoglycemia, and those treated with insulin. However, we note that despite the IDF statement, the proportion of patients with type 2 diabetes who are willing to be tested for OSA might be low (37) and that the effect of continuous positive airway pressure treatment on diabetes-related outcomes from randomized controlled trials remains uncertain or lacking (except with regard to BP) (38–40). Nonetheless, in general population studies, continuous positive airway pressure has been shown to improve quality of life and sleepiness in patients with and without diabetes (38,41).

OSA in patients with type 2 diabetes has been linked to worse glycemic control, higher BP, CVD, and microvascular complications (11,16,17,42,43). In all of these studies, whether the diagnosis of OSA preceded or followed the diagnosis of type 2 diabetes was not clear. In view of our data showing that patients with type 2 diabetes are at increased risk of developing OSA, it would be of interest to examine whether the effect of OSA in these patients differs depending on whether OSA developed before or after the diagnosis of type 2 diabetes.

This study included a large sample size from the THIN database, which is generalizable to the U.K. population. Routinely collected data may be subject to potential biases resulting from incorrect, inconsistent, or incomplete coding of conditions. However, the conditions considered in this study are mostly part of the QOF, and recording quality is therefore expected to be high. The prevalence of major conditions and death rates in THIN, adjusted for patient demographics, are similar to national rates (44).

Ethnicity is not well recorded in THIN; as a result, ethnicity data were not available for all patients in the sample. In addition, adiposity measures such as neck and waist circumference are poorly recorded in primary care and, hence, were not included as covariates in our analysis. It is possible that the observed association between diabetes and OSA was driven by these measures of central obesity, which are common to both diabetes and OSA. Nonetheless, our findings were independent of BMI, which is commonly used in large epidemiological studies and usually correlates with these measures of adiposity.

The prevalence of OSA observed in patients at baseline was lower than the prevalence reported in other studies (12,13,45) because the routinely collected primary care–based OSA diagnosis is subject to underdiagnosis, underrecording, and differences in temporal and interclinic diagnostic criteria. We introduced a 15-month latency period at the start of the study to limit the possibility of silent OSA preceding the diabetes diagnosis being classified as incident OSA. However, we cannot completely rule out the possibility that patients with undiagnosed OSA were included in the study because it was not possible to investigate patients for OSA at baseline. It is also possible that the effect of increased surveillance might have had an effect on the OSA detection rate in patients with diabetes, particularly because symptoms of OSA and diabetes, such as fatigue and sleep disturbance, might overlap. The exploratory analysis for COPD reported above suggests the effect of surveillance bias resulting from increased contact with health care practitioners is likely to be small; however, surveillance bias resulting from differential screening and detection rates in the exposed/unexposed groups cannot be completely ruled out.

Although we cannot determine the methods used to diagnose OSA in our study, patients in the U.K. are usually diagnosed with OSA after an assessment in a sleep clinic/center as described in the National Institute of Clinical Excellence and the British Lung Foundation guidance (46). Nonetheless, the methods (for example, oximetry vs. polygraphy vs. polysomnography) and criteria (oxygen desaturation index vs. apnea-hypopnea index and different cutoffs of the apnea-hypopnea index or oxygen desaturation index) used to diagnose OSA could vary between different sleep centers.

Patients with type 2 diabetes are at increased risk of developing OSA, and the association remained significant after adjusting for potential confounders. In addition to known predictors of OSA, diabetes-related foot disease and insulin treatment were identified as risk factors for OSA in patients with type 2 diabetes. Clinicians should consider testing for OSA in patients with type 2 diabetes, particularly in men, and patients with high BMI, diabetes-related foot disease, CVD, hypertension, depression, and those prescribed insulin because these were shown to be independent risk factors for incident OSA. When taken together with previous evidence, this study indicates that the association between type 2 diabetes and OSA is bidirectional. Further research is required to investigate whether the sequence in which the two diseases develop has an effect on outcomes in patients with type 2 diabetes and OSA.

Funding. A.A.T. is a clinician scientist supported by the National Institute for Health Research (research training fellowship [RTF/01/094] and clinician scientist award [CS-2013-13-029]) in the U.K. The views expressed in this publication are those of the author(s) and not necessarily those of the National Health Service, the National Institute for Health Research, or the Department of Health.

Duality of Interest. A.S. and K.N. received funding from AstraZeneca (RSBD20464). K.N. reports fees from Sanofi and Boehringer Ingelheim outside the submitted work. No other potential conflicts of interest relevant to this article were reported.

The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Author Contributions. A.S., N.J.A., A.T., T.T., W.H., K.A.T., G.N.T., A.A.T., and K.N. reviewed and revised the manuscript and agreed to submission of the final manuscript. A.S., N.J.A., A.T., and A.A.T. wrote the initial draft of the manuscript. A.S., N.J.A., and K.N. designed and performed the analysis. A.A.T and K.N. conceived the idea for the study. K.N. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.