Comments
Abstract
Rationale: Despite emerging evidence that obstructive sleep apnea (OSA) may cause metabolic disturbances independently of other known risk factors, it remains unclear whether OSA is associated with incident diabetes.
Objectives: To evaluate whether risk of incident diabetes was related to the severity and physiologic consequences of OSA.
Methods: A historical cohort study was conducted using clinical and provincial health administrative data. All adults without previous diabetes referred with suspected OSA who underwent a diagnostic sleep study at St. Michael’s Hospital (Toronto, Canada) between 1994 and 2010 were followed through health administrative data until May 2011 to examine the occurrence of diabetes. All OSA-related variables collected from the sleep study were examined as predictors in Cox regression models, controlling for sex, age, body mass index, smoking status, comorbidities, and income.
Measurements and Main Results: Over a median follow-up of 67 months, 1,017 (11.7%) of 8,678 patients developed diabetes, giving a cumulative incidence at 5 years of 9.1% (95% confidence interval, 8.4–9.8%). In fully adjusted models, patients with apnea–hypopnea index (AHI) greater than 30 had a 30% higher hazard of developing diabetes than those with AHI less than 5. Among other OSA-related variables, AHI in rapid eye movement sleep and time spent with oxygen saturation less than 90% were associated with incident diabetes, as were heart rate, neck circumference, and sleep time.
Conclusions: Among people with OSA, and controlling for multiple confounders, initial OSA severity and its physiologic consequences predicted subsequent risk for incident diabetes.
Published reports on the causal relationship between obstructive sleep apnea (OSA) and incident diabetes are very limited. Among only six longitudinal studies published to date, five found a significant association between OSA and incident diabetes. However, these studies were generally small, had few events, did not account for time-to-event in their analyses, and used inconsistent definitions of OSA. There is a need for a larger study with rigorous assessment of both OSA and diabetes, with sufficient follow-up time to allow development of disease.
Based on a large clinical cohort, our study shows that among people with OSA, and controlling for known risk factors for diabetes development, initial OSA severity predicted risk for incident diabetes: in fully adjusted models, patients with apnea-hypopnea index (AHI) greater than 30 had a 30% higher hazard of developing diabetes than those with AHI less than 5. AHI during rapid eye movement sleep and measures of the physiologic consequences of OSA (e.g., oxygen desaturation, sleep deprivation, and sympathetic activation) were also risk factors for diabetes in this population. Risk-stratification of patients with OSA according to these sleep apnea–related predictors may be useful in identifying those most likely to develop diabetes, allowing timely intervention.
Diabetes has been described as a public health epidemic, afflicting 10.8% of women and 11.8% of men in the United States (1). There is emerging evidence that obstructive sleep apnea (OSA), through chronic intermittent hypoxemia, recurrent arousals, and neurohumoral changes, may cause metabolic disturbances including insulin resistance independently of other known risk factors (2–5) and that OSA may represent a therapeutic target in this condition (6).
It remains unclear whether OSA may lead to incident diabetes (3, 7, 8). Among six longitudinal studies, five found a significant association between OSA and incident diabetes (9–13). However, these studies were generally small, had few events, did not account for time-to-event in their analyses (9, 11, 12), and used inconsistent definitions of OSA (e.g., apnea-hypopnea index [AHI] ≥ 5 [9], AHI ≥ 8 [10], oxygen desaturation index > 5 [12, 13]). Furthermore, there has been very limited exploration of the prognostic value of other possibly pathophysiologically relevant OSA-related factors (e.g., arousals, total sleep time [TST]) (10). One large community-based study (the Wisconsin Sleep Cohort) reported an association between OSA and prevalent, but not incident diabetes (14). However, the number of events occurring within follow-up time was small to detect a true relationship (n = 26), too many predictors for that number of events were included in analyses, and the logistic regression used does not take into account the timing of the events.
There is a need for a larger study with rigorous assessment of both OSA and diabetes, with sufficient follow-up time to allow development of disease. We evaluated whether risk of incident diabetes was related to the severity and physiologic consequences of OSA in a large historical cohort of patients studied with in-laboratory polysomnography over more than a decade and whose health information was obtained through provincial health administrative data. Some of the results of these studies have been previously reported in the form of an abstract (15).
We included patients who were referred with suspected OSA and underwent a first diagnostic sleep study at St. Michael’s Hospital (Toronto, Canada) between September 1, 1994 and December 31, 2010. Sleep laboratory clinical data were linked to health administrative data at the Institute for Clinical Evaluative Sciences (ICES, Ontario, Canada) from July 1, 1991 to March 31, 2011. The ethics committees of all institutions involved (St. Michael’s Hospital, ICES, University of Toronto) approved the study. Details on cohort description are provided elsewhere (16).
The St. Michael’s Hospital Sleep Lab database includes a large set of clinical, demographic, and polysomnographic (PSG) variables that have been collected for research purposes since 1991 (see Table E1 in the online supplement). Each patient in the cohort underwent full in-laboratory PSG recording. Disease-specific symptoms and history were collected using standardized questionnaires.
Residents of Ontario have universal public health insurance, the Ontario Health Insurance Plan (OHIP), covering all medically necessary services. All Ontario residents are eligible for OHIP coverage after 3 months of residency in the province (17). Legislation prohibits the private delivery of services covered under OHIP, including laboratory testing. Since 1991, ICES has housed high-quality administrative data on publicly funded services provided, including individual-level information on physician claims, acute care hospitalization, and emergency department visits within Ontario (18). The eligibility of cohort participants for health insurance and their vital status through the follow-up period were assessed using data from the Registered Persons Database. Administrative data regarding claims for continuous positive airway pressure (CPAP) therapy through the Ontario Assistive Devices Program (19) have been available since 2004. A further administrative dataset used for this analysis, the Ontario Diabetes Database (ODD), was developed to establish population-based incidence and prevalence of diabetes in Ontario (20, 21). In addition to the usual ICES data from 1991, the ODD captures hospitalizations during the time period 1988–1990. Details of variables derived from administrative datasets and detailed descriptions of all datasets used are provided in the online supplement (see Tables E2 and E3).
Patients who had undergone a first diagnostic sleep study during the defined study period, and who had a diagnosis of OSA (AHI ≥ 5) or suspected OSA (referred with sleep apnea, but with AHI < 5) were extracted from the Sleep Lab database. Patients were excluded if they (1) underwent split-night, (2) had more than 50% central events or (3) AHI less than 5 and a diagnosis of another sleep disorder, or (4) had prevalent diabetes, from the ODD, at any time between April 1988 and the diagnostic sleep study.
The following variables were derived from clinical data and considered as possible predictors in our statistical models: (1) PSG indices—TST, AHI during TST, rapid eye movement sleep (REM-AHI), and non-REM sleep (non–REM-AHI), arousals index, total number of awakenings, mean SaO2, duration of SaO2 less than 90% (TiSaO2 < 90%), mean heart rate (HR), and the percentage of each sleep stage; (2) clinical symptoms—daytime sleepiness (DS), identified by mean of the Epworth Sleepiness Scale or a positive answer to the question “During the day, do you ever fall asleep unintentionally”?, and self-reported snoring; (3) neck circumference; and (4) self-reported family history of OSA or snoring.
The AHI was defined as the number of apneas and hypopneas per hour of sleep. Hypopnea was consistently defined during the study period as a decrease of more than 50% of the baseline amplitude of breathing for at least 10 seconds; or a clear but smaller decrease in amplitude for at least 10 seconds that is associated with either an SaO2 drop of greater than or equal to 3% or an arousal (22). Patients were classified as not having OSA (AHI < 5), or with mild (AHI of 5–14.9), moderate (AHI of 15–30), or severe (AHI > 30) OSA (23).
The primary outcome was time from the diagnostic PSG to incident diabetes derived from the ODD (20). The ODD uses a validated algorithm that identifies people with diabetes as those having at least one hospitalization record or at least two physician services claims bearing a diagnosis of diabetes within a 2-year period. This algorithm is highly sensitive (86%) and specific (97%) for identifying patients in whom diabetes was recorded in primary care charts; positive predictive value is 80% (20). Use of the first service date was considered as the incident diabetes date. Subjects were followed from their first diagnostic sleep study to the end of March 2011, or the occurrence of a primary outcome or all-cause mortality, whichever occurred first.
The following potential confounders were extracted from the Sleep Lab database: age, sex, body mass index (BMI), waist circumference, and self-reported smoking and alcohol consumption. Comorbidities at baseline (stroke, myocardial infarction, chronic heart failure, hypertension, and the Johns Hopkins’ aggregated diagnosis groups [ADGs]) were identified from administrative data over a 3-year period before the diagnostic sleep study. Comorbidity at baseline based on ADGs was categorized as low (0–5 ADGs), medium (6–10 ADGs), or high (≥11 ADGs) (24). Each patient was assigned to an income quintile using the patient’s postal code.
Descriptive statistics were calculated for relevant data. Crude incidence rates for diabetes per 100 person-years were calculated for the entire sample and by OSA severity (23, 25). In a frail population, death, termed a competing event, may preclude the occurrence of diabetes and lead to overestimation of incidence by the usual Kaplan-Meier method (26). Therefore, we estimated incidence with the cumulative incidence function, which accounts for competing risks. Formal tests for differences in incident diabetes and all-cause mortality between groups were performed using the modified chi-square statistic (27).
We used multivariable Cox regression models to investigate the relationships between OSA-related predictors and incident diabetes, and expressed the results as hazard ratios (see online supplement) (28, 29). AHI was treated as a continuous and categorical variable. We used restricted cubic spline transformations for continuous explanatory variables if nonlinearity was observed, and the resulting standardized hazard ratios compare the 75th and 25th percentiles (28). To confirm findings from traditional multivariable Cox regression model, the Fine and Gray competing-risk regression model was used (30).
For missing variables we used multivariate imputation by chained equations to generate five complete datasets (31) and pooled the coefficients (32). For a unified presentation of all results and figures, the findings shown are for a single imputed dataset. Pooled confidence intervals across imputations for OSA-related variables were at most 2% wider than those presented.
Systematic reviews (8, 33) and expert opinion found age, sex, smoking status, cardiovascular comorbidities, BMI, AHI, TST, and DS to be important for predicting diabetes, so these variables were forced into the models. Although waist circumference is a more accurate measure of obesity than BMI, BMI was chosen to be included in our statistical model because it improved model fit compared with waist circumference and is easier to obtain in routine clinical practice and less affected by measurement error. Other variables were chosen for inclusion if they were selected by backward step-down variable deletion (34) in at least three imputed datasets. We investigated a priori defined interactions between AHI and DS, BMI, age, sex, and cardiovascular disease at baseline (8).
We used the bootstrap for internal validation and overfitting-corrected calibration. Discriminative ability was assessed using Harrell C-index and predictive ability using the model likelihood ratio chi-square statistic (28).
To address the concern that the exact time of the incident diabetes is unknown, we used a binomial regression with the complementary log-log link function, which allows incorporation of different follow-up time for each subject in the model to estimate incidence rate ratio (35, 36).
In the post-2004 cohort with information on CPAP claims, the final model was refitted with the addition of a time-dependent CPAP treatment variable (see Figure E1). To assess the effect of OSA-related predictors on an untreated sample, patients were censored at the time of a CPAP claim.
Additional sensitivity analyses included the following: analyses in which only participants who were eligible for OHIP all of each year (i.e., not out of the province during the year) with at least 5-year look-back window and at least 2-year follow-up; all gave the similar results with the main analyses (data not shown). Finally, the statistical models were refitted on the entire sample including participants who underwent split-night study (see Table E6).
Finally, to assess the sensitivity of results to unmeasured confounders, we used the approach recommended by Lin and coworkers (37). Additional details on the method are provided in the online supplement.
All statistical analyses were performed using R version 2.15.2 (http://www.r-project.org) and SAS version 9.3 (SAS Institute Inc., Cary, NC).
We expected between 162 and 1,038 events based on an anticipated sample of 5,000 persons with an average of 5 years of follow-up and reported rates of incident diabetes from 0.65 to 4.15 per 100 person-years (9–12, 14). That would allow us to examine at least 16 predictors, using the rule of thumb of 10 events per predictor (38).
Between January 1, 1994 and December 31, 2010, a total of 11,596 individuals underwent a first diagnostic sleep study and 10,149 (88%) were linked to administrative datasets (see Figure E2). Patients who were not linked had similar OSA severity and demographic characteristics, but fewer cardiovascular comorbidities and greater DS (16). Our final analyses included 8,678 participants without diabetes at baseline. Table 1 shows baseline characteristics of patients for the entire sample and by OSA severity. The included sample had 62% males, a median age of 48 years, and a median AHI of 15. The amount of missing data ranged from 0.7% (AHI) to 10.1% (TiSaO2 < 90%), 2.4% was missed for BMI and TST, 6.8% for DS, 7.8% for HR, and 8.2% for smoking status (16).
| By OSA Severity for Sample without Diabetes (n = 8,678) | ||||||
|---|---|---|---|---|---|---|
| Variables | With Diabetes (n = 1,471) | Without Diabetes (n = 8,678) | AHI < 5 (n = 1,959) | 5 ≤ AHI < 15 (n = 2,410) | 15 ≤ AHI ≤30 (n = 1,975) | AHI > 30 (n = 2,334) |
| Demographic characteristics | ||||||
| Male | 909 (61.8) | 5,377 (62) | 893 (45.6) | 1,399 (55.6) | 1,238 (62.7) | 1,808 (77.5) |
| Age, yr | 59.0 (50.0–68.0) | 48.0 (38.0–58.0) | 42.0 (33.0–51.0) | 47.0 (38.0–57.0) | 50.0 (41.0–59.0) | 51.0 (43.0–61.0) |
| Clinical symptoms and findings from physical examination | ||||||
| DS, yes* | 663 (45.1) | 2,994 (34.5) | 629 (32.1) | 754 (31.3) | 614 (31.1) | 966 (41.4) |
| ESS total (0–24) | 8.0 (5.0–12.0) | 8.0 (5.0–12) | 8.0 (5.0–12.0) | 8.0 (4.0–12.0) | 8.0 (5.0–12.0) | 8.0 (5.0–12.0) |
| BMI, kg/m2 | 32.0 (28.1–37.6) | 28.4 (25.1–32.7) | 25.8 (22.9–29.6) | 27.8 (24.8–31.5) | 28.8 (25.7–32.8) | 31.1 (27.5–35.7) |
| NC, cm | 41.0 (38.0–44.0) | 39.0 (36.0–42.0) | 37.0 (33.0–39.0) | 38.0 (36.0–41.0) | 40 (37.0–42.0) | 41.0 (39.0–44.0) |
| History | ||||||
| Smoking status, self-reported | ||||||
| Current | 193 (13.1) | 1,646 (19.0) | 377 (19.2) | 476 (19.8) | 342 (17.3) | 435 (18.6) |
| Ex-smoker | 351 (23.9) | 1,518 (17.5) | 263 (13.4) | 373 (15.5) | 362 (18.3) | 501 (21.5) |
| Never | 761 (51.7) | 4,845 (55.8) | 1,181 (60.3) | 1.354 (56.2) | 1,075 (54.4) | 1,207 (51.7) |
| Prior HTN | 1,004 (68.3) | 2,638 (30.4) | 340 (17.4) | 611 (25.4) | 652 (33.0) | 1,019 (43.7) |
| Prior AMI | 159 (10.8) | 241 (2.8) | 26 (1.3) | 50 (2.1) | 58 (2.9) | 106 (4.5) |
| Prior Stroke | 77 (5.2) | 146 (1.7) | 16 (0.8) | 46 (1.9) | 27 (1.4) | 56 (2.4) |
| Prior CHF | 282 (19.2) | 336 (3.9) | 41 (2.1) | 71 (2.9) | 72 (3.6) | 151 (6.5) |
| ADGs | ||||||
| Low (0–5) | 537 (36.5) | 5,112 (58.9) | 1,054 (53.8) | 1,426 (59.2) | 1,148 (58.1) | 1,453 (62.3) |
| Medium (6–10) | 776 (52.8) | 3,161 (36.4) | 775 (39.6) | 885 (36.7) | 686 (34.7) | 789 (33.8) |
| High (≥11) | 158 (10.7) | 405 (4.7) | 130 (6.6) | 99 (4.1) | 78 (3.9) | 92 (3.9) |
| Income status | ||||||
| Q1 (poorest) | 382 (26) | 1,609 (18.5) | 355 (18.1) | 426 (17.7) | 371 (18.8) | 448 (19.2) |
| Q2 | 296 (20.1) | 1,553 (17.9) | 365 (18.6) | 432 (17.9) | 327 (16.6) | 412 (17.7) |
| Q3 | 224 (15.2) | 1,397 (16.1) | 323 (16.5) | 405 (16.8) | 270 (13.7) | 385 (16.5) |
| Q4 | 223 (15.2) | 1,504 (17.3) | 322 (16.4) | 417 (17.3) | 344 (17.4) | 413 (17.7) |
| Q5 (wealthiest) | 333 (22.6) | 2,532 (29.2) | 574 (29.3) | 710 (29.5) | 581 (29.4) | 653 (28.0) |
| PSG indexes | ||||||
| TST, hr | 5.4 (4.4–6.1) | 5.8 (5.0–6.5) | 5.9 (5.0–6.5) | 5.9 (5.1–6.5) | 5.9 (5.0–6.5) | 5.6 (4.7–6.3) |
| AHI, total in TST, events/hr | 25.7 (11–51.7) | 14.7 (5.6–32.0) | 2.0 (0.8–3.5) | 9.3 (7.1–11.9) | 20.9 (17.7–25.1) | 48.9 (37.4–68.4) |
| AHI, total in REM, events/hr | 37.3 (14.1–59.0) | 23.5 (7.7–46.2) | 4.6 (1.4–10.2) | 20.1 (10.6–31.2) | 35.8 (20.0–50.7) | 52.8 (32.6–69.6) |
| Arousals index, total, events/hr | 30.3 (17.7–50.8) | 22.1 (13.5–36.5) | 11.4 (7.8–16.6) | 16.4 (12.1–21.8) | 25.2 (19.4–31.5) | 48.2 (35.8–64.5) |
| AWK in TST, number of events | 28 (20–41) | 24.0 (18.0–34.0) | 21.0 (15.0–27.0) | 23.0 (17.0–31.0) | 25.0 (19.0–34.0) | 32.0 (22.0–45.0) |
| TST90SaO2, min | 4.6 (0.3–35.6) | 0.3 (0–6.5) | 0 (0–0.1) | 0.1 (0–1.5) | 0.9 (0.0–7.2) | 10.2 (1.0–46.1) |
| Mean SaO2, % | 94.1 (92.3–95.5) | 95.0 (93.6–96.1) | 95.9 (94.8–96.8) | 95.2 (94.1–69.3) | 94.9 (93.7–95.9) | 94.0 (92.3–95.2) |
| HR, mean in TST, bpm | 65.9 (58.9–74.5) | 62.4 (56.2–69.2) | 61.8 (55.6–68.6) | 61.7 (55.5–67.9) | 62.2 (55.7–69.0) | 63.6 (57.5–70.8) |
| Incident diabetes | — | 1,017 (11.7) | 166 (8.5) | 253 (10.5) | 216 (10.9) | 367 (15.7) |
| Follow-up time, mo | — | 67.2 (32.6–104.1) | 95.1 (58.4–123.7) | 71 (36.6–105) | 57.2 (28.1–94) | 48.8 (22.3–85.7) |
Over a median follow-up of 67 months, 1,017 (11.7%) participants experienced incident diabetes, giving an incidence rate of 2 per 100 person-years. The potential competing event, death without diagnosed diabetes, occurred in 395 subjects. Cumulative incidence of diabetes at 5 years for the entire sample was 9.1% (95% confidence interval, 8.4–9.8%); for patients with mild OSA it was 7.5% (6.3–8.6%), with moderate OSA 9.9% (8.3–11.4%), and with severe OSA 14.9% (13.2–16.6%). The unadjusted difference in incidence of diabetes was significant (P < 0.0001) between patients with severe OSA and AHI less than 5.
Table 2 shows the HR estimates and model fit statistics for the two classes of models we examined, with AHI as a continuous or categorical variable. In fully adjusted model, severe OSA as defined by AHI was significantly associated with incident diabetes. Patients with severe OSA had a 30% higher hazard of developing diabetes compared with those without OSA, whereas mild and moderate OSA had a 23% higher hazard (Table 2, Figure 1). Among other OSA-related predictors, REM-AHI and TiSaO2 less than 90% were consistently associated with incident diabetes (Table 3), as were DS, HR in sleep, neck circumference, and sleep time (Figure 1; see Table E5). All models were well calibrated (all observed and predicted 5-year survival within 2%) and validated (optimism for R2 for all explored models was ≤0.007).
| OSA-related Predictors | Model 1 | Fully Adjusted Model |
|---|---|---|
| AHI total as a continuous variable | ||
| AHI, total, events/hr (32 vs. 6) | 1.13 (1.06–1.20) | 1.06 (0.99–1.13) |
| LR chi-square, df | 754.98 (18) | 841.32 (24) |
| R2 | 0.10 | 0.11 |
| AHI as a categorical variable (reference group: AHI < 5) | ||
| 5 ≤ AHI < 15 | 1.18 (0.97–1.44) | 1.23 (1.00–1.50) |
| 15 ≤ AHI ≤ 30 | 1.24 (1.01–1.53) | 1.23 (1.00–1.51) |
| AHI > 30 | 1.47 (1.20–1.79) | 1.31 (1.07–1.61) |
| LR chi-square, df | 756.11 (20) | 845.6 (26) |
| R2 | 0.10 | 0.11 |
| Control variables | Sex, age,* BMI,* history of smoking status, prior comorbidities within 3-yr look-back window (HTN, MI, ADG categories), and income status | Model 1 + daytime sleepiness, neck circumference,* heart rate in sleep, and TST |
Figure 1. Results from multivariable Cox regression model presented as standardized hazard ratios (comparing 75th percentile with 25th percentile) with shading representing confidence levels (99%, 95%, 90%, 80%, and 70%). ADG = comorbidity defined using aggregated diagnostic groups as low (1), medium (2), or high (3); AHI = apnea-hypopnea index, events per hour; AMI = acute myocardial infarction; BMI = body mass index, kg/m2; daytime sleepiness = identified by a positive answer to the question “During the day, do you ever fall asleep unintentionally?”; HR = heart rate, bpm; income = income quintiles, ranked from poorest (1) to wealthiest (5); NECK = neck circumference, cm; TST = total sleep time, hours.
[More] [Minimize]| OSA-related Predictors | Model 1 | Fully Adjusted Model |
|---|---|---|
| AHI total in REM as a continuous variable | ||
| REM-AHI, total, events/hr (46 vs. 8) | 1.22 (1.11–1.34) | 1.17 (1.07–1.29) |
| LR chi-square, df | 758.11 (18) | 849.36 (24) |
| R2 | 0.10 | 0.11 |
| Sleep time spent with SaO2 less than 90% as a continuous variable | ||
| TiSaO2 < 90%*, min (6.4 vs. 0) | 1.06 (1.02–1.11) | 1.45 (1.20–1.76) |
| LR chi-square, df | 749 (19) | 853.17 (26) |
| R2 | 0.10 | 0.109 |
| Control variables | Sex, age,* BMI,* history of smoking status, prior comorbidities within 3-yr look-back window (HTN, AMI, ADG categories), and income status | Model 1 + daytime sleepiness, neck circumference,* heart rate in sleep time, TST |
In the Fine and Gray regression, the effects of DS, neck circumference, HR, and OSA severity on incident diabetes had similar hazard ratios to those from the Cox regression and remained significant, whereas the effect of sleep time became nonsignificant.
The effect of AHI on incident diabetes significantly decreased with increased BMI (P = 0.0013) and age (P = 0.0326) (see Table E5).
Among 3,931 subjects who underwent a diagnostic PSG between 2004 and 2010, a total of 611 (15.5%) submitted a CPAP claim. Among 267 (6.8%) patients who experienced incident diabetes, 66 claimed CPAP before the incident date (24.7%) and 7 after (2.6%); among the other 3,664 subjects, 538 (14.7%) claimed for CPAP treatment. A claim for CPAP treatment had a nonsignificant effect in fully adjusted models on the risk of diabetes (P values > 0.2). When models were refitted on an untreated sample, all predictors except DS remained significantly associated with the outcome.
After accounting for follow-up time in the binomial regression model, AHI, REM-AHI, and TiSaO2 less than 90% remained significant, as did sleep time, DS, HR, and neck circumference.
In a large clinical cohort without diabetes at baseline, 11.7% of subjects experienced incident diabetes over a median 67 months of follow-up. The multivariable Cox regression models identified that OSA severity, expressed as AHI, was independently and significantly associated with incident diabetes. In addition, the OSA-related factors REM-AHI and TiSaO2 less than 90% were significant predictors, as were shorter TST, higher mean HR, greater neck circumference, and the presence of DS. In an untreated subsample, all predictors except DS remained significantly associated with the outcome. The effects of predictors were consistent in a model adjusting for the competing risk of all-cause mortality and in a binomial regression accounting for imprecision in the date of diagnosis of diabetes. The present study agrees with the much smaller single study that looked at time-to-diagnosis of diabetes by Botros and coworkers (10), which found an independent association between OSA and incident diabetes after adjusting for multiple confounding factors.
Our study addresses limitations of previously published observational studies of OSA and diabetes. Because of its larger size (>8,500 patients) and longer period of complete follow-up (>10 yr), our study was able to analyze a number of events that is more than an order of magnitude larger than occurred in any previous study, including the Wisconsin Cohort Study. Furthermore, this allowed assessment of many OSA-related factors beyond AHI and adequate control for numerous potential confounders impossible in a smaller study. Clinical data were consistently collected and used the same PSG scoring criteria over time. We included patients with a wider range of OSA severity than observed in community-based studies and a relatively large number of females. We used validated algorithms to define prevalent diabetes and comorbidities at baseline, and used rigorous methods for missing data, model selection, calibration, and validation.
Our findings (that longer TiSaO2 < 90%, shorter sleep time, and higher HR increase the risk of diabetes) are consistent with proposed pathophysiologic mechanisms (oxidative stress caused by intermittent hypoxemia, sleep deprivation or sleep fragmentation, and sympathetic activation) whereby OSA may lead to diabetes. Severity of hypoxemia has been found to be associated with glucose intolerance and insulin resistance (39, 40). Sleep deprivation may act through sympathetic activation and subsequent alterations in hypothalamic-pituitary-adrenal axis (4, 41). Elevated resting HR, an indicator of sympathetic nerve activity, has been shown to be associated with incident diabetes through possible relation between sympathetic activity and insulin resistance (42, 43).
Similarly, neck circumference, a strong clinical predictor of OSA (44) and a significant independent predictor of incident diabetes in our study, has also been associated with impaired glucose homeostasis and cardiometabolic syndrome (45, 46).
We found that in addition to overall AHI during TST, REM-AHI was also an independent predictor of diabetes. Compared with non-REM sleep, REM sleep has been shown to be associated with greater sympathetic activity and respiratory and cardiovascular instability. Apneas during REM sleep lead to greater degrees of hypoxemia and sympathetic activity compared with events in non-REM sleep (47). In a cross-sectional study on a predominantly African American and Hispanic cohort, REM-AHI, but not overall AHI, was significantly and independently associated with diabetes (48). Similarly, we found that REM-AHI was significantly associated with incident diabetes and had a larger effect than overall AHI. The clinical importance of AHI in REM may have significant implications for clinical practice (47).
The decreased effect of AHI with age on incident diabetes found in our study has been observed previously for the relationship between OSA and mortality (49). Protective adaptive physiologic change through longstanding mild chronic intermittent hypoxia is one of the possible explanations (50). A similar effect on the association between AHI and incident diabetes was found for BMI: the effect of AHI decreased with increasing BMI. It is possible that very obese individuals are already at such high risk for developing diabetes that OSA confers little incremental risk.
There are several limitations that should be considered in the interpretation of our findings. As with any observational study, some methodologic issues are related to availability of data. Some important confounders (e.g., family history of diabetes, race) were not available. The generalizability can be affected by the single center design of our study. Also using the ODD to derive incident diabetes, we were unable to distinguish between type 1 and type 2 diabetes. However, we expect most events to be type 2 because of the age of our cohort. Although validated algorithms were used to define prevalent diabetes and prior comorbidities from health administrative data, these algorithms are characterized by certain sensitivity and specificity resulting in possible misclassification of subjects. If differential, bias could go in either direction, whereas if nondifferential (misclassified randomly and independently of disease state), the estimated effect of OSA severity on incident diabetes is more likely to fall below the true value (51, 52). Because of universal health care in Ontario, this measurement error (undiagnosed cases of diabetes in the cohort) was likely independent of the exposure (severe OSA), so we may have underestimated the true effects. With respect to defining incident diabetes cases using the ODD, a small proportion of prevalent cases may be misclassified as incident cases. This may occur when disease was not captured in the administrative health data within observational period before baseline for patients with true diabetes. In addition, the date of diagnosed diabetes in administrative data is not the exact time of diabetes development: it could have occurred any time before this date. Using the complementary log-log link regression we tried to address this limitation and have not revealed any important differences compared with the Cox regression model.
Finally, patients with more severe OSA may have more contact with the healthcare system and may be more likely to be tested for diabetes and consequently diagnosed with diabetes. We addressed this issue by adjusting our models for age, sex, BMI, and baseline comorbidities, known predictors of healthcare use for patients with OSA (53). Among variables tested in our models, older age, hypertension, and higher income have been also shown to be associated with a higher likelihood of having a glucose test (54). Furthermore, we assessed the sensitivity of the results to unmeasured confounders and found that only fairly strong confounding reduced the hazard ratio of 1.31 sufficiently that it was no longer statistically significant (see Table E4 and Figure E3). In particular, the odds ratio between the confounder and OSA severity needed to be 2.5–3. Although an unmeasured confounder could be any unmeasured feature of the patient, we were most concerned about screening for diabetes. Because it is implausible that screening occurs 2.5–3 times more often in those with AHI greater than 30 compared with those with AHI less than 5, when we have already accounted for age, sex, prior comorbidities, and income status, we believe that confounding by diabetes screening is not solely responsible for the observed hazard ratio of 1.31. The true hazard ratio may be lower than 1.31, but remains statistically significantly elevated for reasonable assumptions about unmeasured confounding.
The nonsignificant effect of treatment that we found in the post-2004 cohort could be explained by lack of information about CPAP adherence, treatment approaches other than CPAP, and reduced sample size for this analysis. Also, patients would have been suffering from physiologic consequences of OSA for many years before starting treatment that could have increased their risk of developing adverse long-term consequences. Nevertheless, because a treatment effect may attenuate a possible association between OSA and incident diabetes, we conducted an additional analysis on untreated subsample only. This subsample analysis replicates the results obtained on the entire cohort. Thus, the nonsignificant association between CPAP claims and the outcome of interest should not be interpreted as a lack of efficacy of CPAP treatment in preventing diabetes because our study was not designed to address this question.
Our study shows that among people with OSA, and controlling for multiple confounders, initial OSA severity predicted risk for incident diabetes. AHI during REM sleep and measures of the physiologic consequences of OSA (e.g., oxygen desaturation, sleep deprivation, and sympathetic activation) were also risk factors for diabetes in this population. Risk-stratification of patients with OSA according to these OSA parameters may be useful in identifying those most likely to develop diabetes, allowing timely intervention.
| 1. | Centers for Disease Control and Prevention. National diabetes fact sheet: national estimates and general information on diabetes and prediabetes in the United States, 2011. Atlanta, GA: US Department of Health and Human Services, Centers for Disease Control and Prevention; 2011. |
| 2. | Tasali E, Ip MS. Obstructive sleep apnea and metabolic syndrome: alterations in glucose metabolism and inflammation. Proc Am Thorac Soc 2008;5:207–217. |
| 3. | Clarenbach CF, West SD, Kohler M. Is obstructive sleep apnea a risk factor for diabetes? Discov Med 2011;12:17–24. |
| 4. | Punjabi NM, Polotsky VY. Disorders of glucose metabolism in sleep apnea. J Appl Physiol (1985) 2005;99:1998–2007. |
| 5. | Sharma S, Kavuru M. Sleep and metabolism: an overview. Int J Endocrinol 2010;2010:270832. |
| 6. | Sharma SK, Agrawal S, Damodaran D, Sreenivas V, Kadhiravan T, Lakshmy R, Jagia P, Kumar A. CPAP for the metabolic syndrome in patients with obstructive sleep apnea. N Engl J Med 2011;365:2277–2286. |
| 7. | Pamidi S, Tasali E. Obstructive sleep apnea and type 2 diabetes: is there a link? Front Neurol 2012;3:126. |
| 8. | Kendzerska T, Mollayeva T, Gershon AS, Leung RS, Hawker G, Tomlinson G. Untreated obstructive sleep apnea and the risk for serious long-term adverse outcomes: a systematic review. Sleep Med Rev 2014;18:49–59. |
| 9. | Marshall NS, Wong KK, Phillips CL, Liu PY, Knuiman MW, Grunstein RR. Is sleep apnea an independent risk factor for prevalent and incident diabetes in the Busselton Health Study? J Clin Sleep Med 2009;5:15–20. |
| 10. | Botros N, Concato J, Mohsenin V, Selim B, Doctor K, Yaggi HK. Obstructive sleep apnea as a risk factor for type 2 diabetes. Am J Med 2009;122:1122–1127. |
| 11. | Celen YT, Hedner J, Carlson J, Peker Y. Impact of gender on incident diabetes mellitus in obstructive sleep apnea: a 16-year follow-up. J Clin Sleep Med 2010;6:244–250. |
| 12. | Lindberg E, Theorell-Haglöw J, Svensson M, Gislason T, Berne C, Janson C. Sleep apnea and glucose metabolism: a long-term follow-up in a community-based sample. Chest 2012;142:935–942. |
| 13. | Muraki I, Tanigawa T, Yamagishi K, Sakurai S, Ohira T, Imano H, Kitamura A, Kiyama M, Sato S, Shimamoto T, et al.; CIRCS Investigators. Nocturnal intermittent hypoxia and the development of type 2 diabetes: the Circulatory Risk in Communities Study (CIRCS). Diabetologia 2010;53:481–488. |
| 14. | Reichmuth KJ, Austin D, Skatrud JB, Young T. Association of sleep apnea and type II diabetes: a population-based study. Am J Respir Crit Care Med 2005;172:1590–1595. |
| 15. | Kendzerska T, Leung, R., Gershon, A., Hawker, G., Tomlinson, G. Obstructive sleep apnea and incident diabetes. Sleep2013;36;A123. |
| 16. | Kendzerska T, Gershon AS, Hawker G, Leung RS, Tomlinson G. Obstructive sleep apnea and risk of cardiovascular events and all-cause mortality: a decade-long historical cohort study. PLoS Med 2014;11:e1001599. |
| 17. | Ministry programs. Ontario Health Insurance Plan [2014 March 22]. Available from: http://www.health.gov.on.ca/en/public/programs/ohip/ |
| 18. | Improving health care data in Ontario. ICES investigative report Toronto: Institute for Clinical Evaluative Sciences; 2005. Available from: http://www.ices.on.ca/flip-publication/improving-health-care-data/index.html#3/z |
| 19. | Policies and procedures manual for the assistive devices program. Assistive Devices Program Ministry of Health and Long-Term Care. 2012. Available from: http://www.health.gov.on.ca/en/pro/programs/adp/update_vendor_agreements/docs/pp_adp_manual.pdf |
| 20. | Hux JE, Ivis F, Flintoft V, Bica A. Diabetes in Ontario: determination of prevalence and incidence using a validated administrative data algorithm. Diabetes Care 2002;25:512–516. |
| 21. | Lipscombe LL, Hux JE. Trends in diabetes prevalence, incidence, and mortality in Ontario, Canada 1995-2005: a population-based study. Lancet 2007;369:750–756. |
| 22. | Sleep-related breathing disorders in adults: recommendations for syndrome definition and measurement techniques in clinical research. The Report of an American Academy of Sleep Medicine Task Force. Sleep 1999;22:667–689. |
| 23. | Fleetham J, Ayas N, Bradley D, Ferguson K, Fitzpatrick M, George C, Hanly P, Hill F, Kimoff J, Kryger M, et al. Canadian Thoracic Society guidelines: diagnosis and treatment of sleep disordered breathing in adults. Can Respir J 2006;13:387–392. |
| 24. | Starfield B, Chang HY, Lemke KW, Weiner JP. Ambulatory specialist use by nonhospitalized patients in US health plans: correlates and consequences. J Ambul Care Manage 2009;32:216–225. |
| 25. | Epstein LJ, Kristo D, Strollo PJ Jr, Friedman N, Malhotra A, Patil SP, Ramar K, Rogers R, Schwab RJ, Weaver EM, et al.; Adult Obstructive Sleep Apnea Task Force of the American Academy of Sleep Medicine. Clinical guideline for the evaluation, management and long-term care of obstructive sleep apnea in adults. J Clin Sleep Med 2009;5:263–276. |
| 26. | Wolbers M, Koller MT, Witteman JC, Steyerberg EW. Prognostic models with competing risks: methods and application to coronary risk prediction. Epidemiology 2009;20:555–561. |
| 27. | Gray R. A class of k-sample tests for comparing the cumulative incidence of a competing risk. Ann Stat 1988;16:1141–1154. |
| 28. | Harrell FE. Regression modeling strategies: with applications to linear models, logistic regression, and survival analysis. New York: Springer; 2001. |
| 29. | Grambsch P, Therneau T. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika 1994;81:515–526. |
| 30. | Fine JP, Gray RJ. A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc 1999;94:496–509. |
| 31. | Azur MJ, Stuart EA, Frangakis C, Leaf PJ. Multiple imputation by chained equations: what is it and how does it work? Int J Methods Psychiatr Res 2011;20:40–49. |
| 32. | Rubin DB. Multiple imputation for nonresponse in surveys. New York: Wiley; 1987. |
| 33. | Collins GS, Mallett S, Omar O, Yu LM. Developing risk prediction models for type 2 diabetes: a systematic review of methodology and reporting. BMC Med 2011;9:103. |
| 34. | Atkinson AC. A note on the generalized information criterion for choice of a model. Biometrika 1980;67:413–418. |
| 35. | Collett D. Modelling binary data. London: Chapman & Hall; 1991. |
| 36. | Martuzzi M, Elliott P. Estimating the incidence rate ratio in cross-sectional studies using a simple alternative to logistic regression. Ann Epidemiol 1998;8:52–55. |
| 37. | Lin DY, Psaty BM, Kronmal RA. Assessing the sensitivity of regression results to unmeasured confounders in observational studies. Biometrics 1998;54:948–963. |
| 38. | Peduzzi P, Concato J, Feinstein AR, Holford TR. Importance of events per independent variable in proportional hazards regression analysis. II. Accuracy and precision of regression estimates. J Clin Epidemiol 1995;48:1503–1510. |
| 39. | Punjabi NM, Shahar E, Redline S, Gottlieb DJ, Givelber R, Resnick HE; Sleep Heart Health Study Investigators. Sleep-disordered breathing, glucose intolerance, and insulin resistance: the Sleep Heart Health Study. Am J Epidemiol 2004;160:521–530. |
| 40. | Redline S, Storfer-Isser A, Rosen CL, Johnson NL, Kirchner HL, Emancipator J, Kibler AM. Association between metabolic syndrome and sleep-disordered breathing in adolescents. Am J Respir Crit Care Med 2007;176:401–408. |
| 41. | Balbo M, Leproult R, Van Cauter E. Impact of sleep and its disturbances on hypothalamo-pituitary-adrenal axis activity. Int J Endocrinol 2010;2010:759234. |
| 42. | Carnethon MR, Yan L, Greenland P, Garside DB, Dyer AR, Metzger B, Daviglus ML. Resting heart rate in middle age and diabetes development in older age. Diabetes Care 2008;31:335–339. |
| 43. | Bemelmans RH, Wassink AM, van der Graaf Y, Nathoe HM, Vernooij JW, Spiering W, Visseren FL. Risk of elevated resting heart rate on the development of type 2 diabetes in patients with clinically manifest vascular diseases. Eur J Endocrinol 2012;166:717–725. |
| 44. | Davies RJ, Stradling JR. The relationship between neck circumference, radiographic pharyngeal anatomy, and the obstructive sleep apnoea syndrome. Eur Respir J 1990;3:509–514. |
| 45. | Fitch KV, Stanley TL, Looby SE, Rope AM, Grinspoon SK. Relationship between neck circumference and cardiometabolic parameters in HIV-infected and non-HIV-infected adults. Diabetes Care 2011;34:1026–1031. |
| 46. | Laakso M, Matilainen V, Keinanen-Kiukaanniemi S. Association of neck circumference with insulin resistance-related factors. Int J Obes Rel Metabolic Disord 2002;26:873–875. |
| 47. | Mokhlesi B, Punjabi NM. “REM-related” obstructive sleep apnea: an epiphenomenon or a clinically important entity? Sleep 2012;35:5–7. |
| 48. | Mahmood K, Akhter N, Eldeirawi K, Onal E, Christman JW, Carley DW, Herdegen JJ. Prevalence of type 2 diabetes in patients with obstructive sleep apnea in a multi-ethnic sample. J Clin Sleep Med 2009;5:215–221. |
| 49. | Lavie P, Lavie L, Herer P. All-cause mortality in males with sleep apnoea syndrome: Declining mortality rates with age. Eur Respir J 2005;25:514–520. |
| 50. | Lavie P, Lavie L. Unexpected survival advantage in elderly people with moderate sleep apnoea. J Sleep Res 2009;18:397–403. |
| 51. | Brenner H, Gefeller O. Use of the positive predictive value to correct for disease misclassification in epidemiologic studies. Am J Epidemiol 1993;138:1007–1015. |
| 52. | Wacholder S, Hartge P, Lubin JH, Dosemeci M. Non-differential misclassification and bias towards the null: a clarification. Occup Environ Med 1995;52:557–558. |
| 53. | Albarrak M, Banno K, Sabbagh AA, Delaive K, Walld R, Manfreda J, Kryger MH. Utilization of healthcare resources in obstructive sleep apnea syndrome: a 5-year follow-up study in men using CPAP. Sleep 2005;28:1306–1311. |
| 54. | Wilson SE, Rosella LC, Lipscombe LL, Manuel DG. The effectiveness and efficiency of diabetes screening in Ontario, Canada: a population-based cohort study. BMC Public Health 2010;10:506. |
Supported by Canadian Institutes of Health Research doctoral research award (T.K.); the ResMed research foundation; the F.M. Hill Chair in Academic Women’s Medicine (G.H.); Physicians’ Services Incorporated Foundation Graham Farquharson Knowledge Translation Fellowship (A.S.G.); and the Institute for Clinical Evaluative Sciences, which is funded by an annual grant from the Ontario Ministry of Health and Long-Term Care.
Author Contributions: T.K., literature search; study conception and design; ethics boards’ application; obtaining administrative data; cleaning, analyses, and interpretation of data; drafting of the manuscript. G.T., study design, data interpretation, drafting of the manuscript, critical revision, supervision of manuscript writing, ethics boards’ application, and data analyses. R.S.L., study design, data interpretation, drafting of the manuscript, critical revision, supervision of manuscript writing, study conception, ethics boards’ application, is an owner of the sleep portion of the Chest Dataset from which the study sample was extracted, and gave final approval of the submitted manuscript. A.S.G., study design, data interpretation, drafting of the manuscript, critical revision, supervision of manuscript writing, ethics boards’ application, obtaining administrative data, and data analyses. G.H., study design, data interpretation, drafting of the manuscript, critical revision, supervision of manuscript writing.
This article has an online supplement, which is accessible from this issue's table of contents at www.atsjournals.org
Originally Published in Press as DOI: 10.1164/rccm.201312-2209OC on June 4, 2014
Author disclosures are available with the text of this article at www.atsjournals.org.
