Study design and participants
The Kailuan study was a prospective cohort study involving 101,510 participants (men: 81,110; women: 20,400, aged 18–98 years) in the Kailuan community from June 2006 to October 2007. Of the 101,510 people who participated in the survey, 34,260 adults were included in the analysis after excluding 62,200 subjects who had a history of dyslipidemia, 1436 subjects with missing information about sleep duration or serum lipid profiles, and 3614 subjects without the data of 2008-2009, 2010-2011,2012-2013, or 2014-2015 follow-up visits(Figure1). The health interview survey was performed using self-administered structured questionnaires to obtain information on sociodemographic characteristics, health status, and health behaviors. Before the study, all doctors and nurses received rigorous unified training.
Assessment of sleep duration
Sleep duration data were collected through a self-reported answer to the question “How many hours of sleep have you had on an average at night in the preceding 3 months?” We divided sleep durations into five groups according to the responses: ≤5 hours, 6 hours, 7 hours, 8 hours, and ≥9 hours. Additionally, participants were asked to answer “yes” or “no” to the question “Do you generally snore when you sleep?”[24, 25].
Assessment of potential covariates
The clinical examination consisted of a medical history, a physical examination, anthropometric measurements, and self-administered questionnaires on lifestyle characteristics [24, 26], such as sleep duration,
regular leisure-time physical activity, smoking habits, and daily
alcohol consumption. Body mass index (BMI) was calculated as the weight (kg) divided by the square of height (meters2). Blood pressure was measured three times using a standardized sphygmomanometer while the participant was in a seated position. We used the average of the three measurements.
We measured TG, TC, LDL-C, HDL-C, fasting blood glucose (FBG), and high-sensitivity C-reactive protein (h-CRP) levels. All blood samples were analyzed by using a Hitachi 747 autoanalyzer (Hitachi; Tokyo, Japan). Diabetes was defined as having a history of diabetes, the use of glucose-lowering agents, or a fasting blood glucose ≥7 mmol/l. Hypertension was defined as a systolic blood pressure ≥140 mmHg, a diastolic blood pressure ≥90 mmHg, a history of hypertension and/or the use of antihypertensive agents.
Diagnosis of each lipid profile abnormality
Each lipid profile abnormality was defined according to the Chinese guidelines on the prevention and treatment of dyslipidemia in adults (2016) . Specifically, abnormal TC was defined as TC≥5.2 mmol/L, abnormal TG were defined as TG ≥1.70 mmol/L, abnormal LDL-C was defined as LDL-C ≥3.4 mmol/L, and abnormal HDL-C was defined as HDL-C <1.0 mmol/L.
We described continuous variables by their means ± standard deviations and compared groups using a one-way analysis of variance (ANOVA). Categorical variables were described as percentages and were compared using the Chi-square test. The Cox proportional hazards model was used to estimate the hazard ratio (HR) for the incidence of each lipid profile abnormality in relation to sleep duration and baseline covariates. The follow-up time was calculated from the 2006 interview to the date when lipid profile abnormalities was detected, date of death, or date of the last attended interview in this analysis, whichever came first. For all models, the adequacy of the Cox proportional hazards model was assessed. Model 1 was adjusted for age and sex. Model 2 further adjusted for the level of education, smoking, alcohol consumption, physical activity, and snoring. Model 3 further adjusted for hypertension, diabetes mellitus, BMI, and h-CRP. Because the transient changes in lipid profile (e.g., reversion from dyslipidemia to normal blood lipid) can impact baseline sleep duration and future dyslipidemia risk, we repeated our analysis after excluding individuals with these conditions(Table3- Sensitivity analysis).To investigate whether age/sex acted as an effect modifier in the relationship between sleep duration and dyslipidemia, we ran a regression model with an interaction term between age/sex and sleep duration. Because 11 hospitals participated in the study, we used a Cox proportional hazards model with a sandwich covariance matrix as a random effect to account for the potential confounding effect of multiple hospitals participating in the study. The statistical analysis was performed using SAS 9.4. All statistical tests were two-sided, and the significance level was set at 0.05.