Increased risk of obstructive sleep apnoea in women with polycystic ovary syndrome: a population-based cohort study

Objective Obesity is very common in patients with obstructive sleep apnoea (OSA) and polycystic ovary syndrome (PCOS). Longitudinal studies assessing OSA risk in PCOS and examining the role of obesity are lacking. Our objective was to assess the risk of OSA in women with vs without PCOS and to examine the role of obesity in the observed findings. Design Population-based retrospective cohort study utilizing The Health Improvement Network (THIN), UK. Methods 76 978 women with PCOS and 143 077 age-, BMI- and location-matched women without PCOS between January 2000 and May 2017 were identified. Hazard ratio (HR) for OSA among women with and without PCOS were calculated after controlling for confounding variables using multivariate Cox models. Results Median patient age was 30 (IQR: 25–35) years; median follow-up was 3.5 (IQR: 1.4–7.1) years. We found 298 OSA cases in PCOS women vs 222 in controls, with incidence rates for OSA of 8.1 and 3.3 per 10 000 person years, respectively. Women with PCOS were at increased risk of developing OSA (adjusted HR = 2.26, 95% CI: 1.89–2.69, P < 0.001), with similar HRs for normal weight, overweight and obese PCOS women. Conclusions Women with PCOS are at increased risk of developing OSA compared to control women irrespective of obesity. Considering the significant metabolic morbidity associated with OSA, clinicians should have a low threshold to test for OSA in women with PCOS. Whether OSA treatment has an impact on PCOS symptoms and outcomes needs to be examined.


Study population
The open cohort extended from the 1 st January 2000 (study start date) to the 15 th May 2017 (study end date). All individuals were required to be registered at their practice at least for a year before entry into the study in order to ascertain documentation of concomitant diseases.
Their practice was also required to have been using their computer system (Vision) for at least one year prior to the index date (the date at which each participant joined the cohort) and have an acceptable mortality reporting in order to ensure good data quality [9].
All women who fulfilled the above-mentioned criteria and who were aged 18-50 years at the index date and had a documentation of PCOS at any time during the study period were included in the exposed group. Patients with any documentation of OSA prior to the index date were excluded. Women without documented PCOS at any time during the study period were included in the unexposed (control) arm. The index date was defined as the date of first documentation of PCOS for newly diagnosed cases and from the date patient became eligible if the first documentation of PCOS was prior to the eligibility date (for existing cases) (see Figure E1 in the online data supplement.).
Each exposed patient was randomly matched to 2 unexposed patients (1:2 ratio) for general practice, age at index date (± one year) and BMI (± 2 kg/m 2 ). Matching variables were chosen based on their strong association with PCOS and OSA [8] [10] When more than 2 participants in the unexposed cohort were available to be matched with a participant in the exposed cohort, two were randomly selected.
To minimize the immortal time bias, each randomly matched eligible unexposed patient was assigned the same index date as their corresponding exposed patient [11]. Follow up end date (exit date) was determined from the earliest occurrence of: the first documentation of OSA, transfer to another practice, death or study-end.

Selection of Read Codes and PCOS definition
Read Codes to define PCOS, OSA and covariates (see Tables E1 to E17 in the online data supplement.) were compiled using a methodical Read Code search strategy (see Panel E1 in the online data supplement.) [12] including codes used in other studies [7] [8]. Since there is a possibility of misclassification between PCOS and Polycystic ovaries (PCO) due to the resemblance of codes during data entry, they have been combined in prevalence studies using general practice electronic databases [7].

Statistical analysis
Potential confounders and covariates were chosen based on biological plausibility and links to the exposure and outcome of interest (PCOS and OSA respectively). These included: age, Townsend social deprivation index [8], BMI, smoking status, diabetes mellitus, impaired glucose regulation [14] hypertension [15], hypothyroidism, antiandrogen medication, and metformin. Baseline characteristics were summarized considering the distribution of data and without statistical testing to comply with the statistical guidelines [16] [17] . The critical value for statistical significance was set at 5% and therefore 95% confidence intervals were used in the population estimate of hazard ratios. Further analysis included calculating the number and percentage of incident cases, person years, and incidence rates. Unadjusted and adjusted hazard ratios were estimated using Cox regression models. The initial adjusted model included age, Townsend score, BMI, diabetes or IGR and hypothyroidism at baseline as covariates. Sensitivity analysis was carried out to assess selection bias due to case definition (PCOS and PCO vs PCOS only) and survival bias due to inclusion of prevalent cases (who had the documentation of PCOS prior to becoming eligible for the study) [18].
Another model confined to PCOS cases (excluding controls) was used to assess the association of PCOS with OSA while considering PCOS phenotypes and antiandrogen medication as covariates. STATA MP version 14.2 was used for data cleaning and analysis [19] 21.