Study design and database
The Korean National Health Insurance System (NHIS) uses billing records of healthcare providers to collect data on age, sex, demographic variables, treatment records, general health examinations, lifestyle, and behavior of patients. Healthcare providers are advised to perform standardized medical examinations every 1 or 2 years. Enrollment in the NHIS is compulsory for all of the > 50 million Korean residents.29,30
Study population
We collected data from 19,459,018 participants in the NHIS who underwent medical checkups between 2012 and 2013 (index year), and had been enrolled in NHIS for at least 5 years before the checkup (i.e., washout period: 2007–2011). We selected 5,632,394 participants who underwent a medical checkup in the index year and ≥ 2 checkups during the preceding 3 years. After exclusion of participants of age < 20 years (n = 435) or missing data (n = 34,810), data from 5,587,754 participants was included in this study (Fig. 1). Ethical approval was obtained from the Institutional Review Board of Uijeongbu St. Mary’s Hospital, Catholic University of Korea (UC18ZESI0094). Informed consent was not required as the data used for this study were anonymized. All methods were carried out in accordance with relevant guidelines and regulations. The need of the written informed consent has waived by the ethics committee of institutional review board of the Korean National Institute for Bioethics Policy and the Catholic University of Korea Institutional Review Board.
Measurements and definitions
Body mass index (BMI) was calculated by dividing the weight (kg) by the square of height (m2), and obesity was defined as BMI ≥ 25 kg/m2.31 “Regular physical activity” was defined as ≥ 20 min of vigorous physical activity ≥ 3 times/week or ≥ 30 min of moderate-intensity physical activity ≥ 5 times/week.
Household income was dichotomized at 25% of monthly income.32 Diabetes mellitus (DM) was recorded as being present if the fasting blood glucose level was ≥ 126 mg/dL or there was ≥ 1 claim/year for International Classification of Disease, tenth revision (ICD-10) codes E10–14 and ≥ 1 claim/year for antidiabetic drug prescriptions. Hypertension was recorded as being present if the blood pressure was ≥ 140/90 mm Hg or there was ≥ 1 claim/year for ICD-10 codes I10–15 and ≥ 1 claim/year for antihypertensive drug prescriptions. Dyslipidemia was recorded as being present if the total cholesterol was ≥ 240 mg/dL or there was ≥ 1 claim/year for ICD-10 code E78 and ≥ 1 claim/year for lipid-lowering drug prescriptions. Self-administered questionnaires were used to document social behaviors, including smoking, alcohol use, and physical activity.
Definition of HDL-C variability
HDL-C variability was calculated using HDL-C levels measured at two different health screenings. HDL-C variability was measured with the coefficient of variation (CV), variability independent of the mean (VIM), and average real variability (ARV). CV was calculated as (standard deviation [SD]/mean) × 100. VIM was calculated as (SD/meanβ) × 100, where β was the regression coefficient.33,34 ARV was the average absolute difference between consecutive HDL-C levels.35
Definition of low mean HDL-C and high HDL-C variability
Mean HDL-C values differ between men and women; therefore, we used sex-specific cutoff values (Table S1). Participants with HDL-C levels in the lowest quartile (quartile 1) were included in the low-mean HDL-C group, and those with HDL-C levels in the remaining three quartiles (quartiles 2–4) were included in the high-mean HDL-C group. Participants with HDL-C variability in the highest quartile (quartile 4) were included in the high-variability group, and those with HDL-C variability in the remaining three quartiles (quartiles 1–3) were included in the low-variability group.
Study outcomes and follow-up
The end point of this study was a diagnosis of BD, defined as documentation of ICD-10 codes M35.2 or V139 (rare intractable diseases). The study participants were followed up from baseline to BD diagnosis or until December 31, 2016, whichever came earlier. The median follow-up duration was 4.22 (4.01–4.55) years.
Statistical analysis
Baseline demographics are presented as mean ± SD, median (interquartile range 25–75%), or n (%). Participants were grouped using HDL-C mean and CV quartiles. Hazard ratios (HRs) and 95% confidence interval (CI) values were calculated using the Cox proportional-hazards model. The HRs (95% CIs) of the low-mean HDL-C and high-variability groups were compared to those of the high-mean HDL-C and low-variability groups, respectively. Kaplan–Meier estimates were used to calculate the cumulative HRs for quartiles of HDL-C mean and variability as well as groups of combinations of mean and variability. These HRs were used in the Schoenfeld residuals test to evaluate the proportional hazards function. There was no significant departure from proportionality of hazards over time. A proportional-hazards model was applied after adjusting for age, sex, BMI, alcohol use, smoking, exercise, income, DM, and hypertension. We used stratified analysis and a likelihood-ratio test to evaluate potential modifier effects of age, sex, obesity, DM, hypertension, malignancy, and use of lipid-lowering agents. Statistical analyses were performed using SAS software (version 9.4; SAS Institute Inc., Cary, NC, USA), and p-values ≤ 0.05 were considered to indicate significance.