Data sources
We analyzed National Health Insurance Service-National Sample Cohort 2.0 (NHIS-NSC 2.0) data set. The Korean government has an obligatory public health insurance program, National Insurance Health Service (NIHS), including more than 97% of Korean people are affiliated. Total population (n=48,222,537) were classified into 2,142 classes according to their age, sex, area, eligibility status, and income level. Then, Korean government randomly selected 2.1% of them from each stratum (n=1,021,208) from 2006 and made NHIS-NSC 2.0. It also included retrospective and prospective follow-up data which was collected from 2002 to 2015.[13] Because the NHIS covers about 97% of the total population of Korea, the NHIS-NSC 2.0 cohort is expected to represent the entire Korean population.
The cohort includes 4 datasets: the dataset of the sociodemographic information; the dataset of medical claims including information on the diagnosis based on the 10th revision of the International Classification of Disease (ICD-10) codes, admission, and treatment; the dataset of the National Health Screening of the cohort members; and the dataset of the medical institutions. The Korean government recommends the entire Korean adults to take the National Health check-up biennially including questionnaires on medical history and health-related behaviors including smoking status and alcohol consumption, chest X-ray, physical examinations and blood tests. About 72.1% of eligible population had National Health screening programs according to the 2013 NHIS statistics [13]. The cohort also includes mortality data from the death registration database of the Statistics Korea, a central government organization for statistics.
The NHIS-NSC 2.0 is open to any researchers if the NHIS review committee approves study protocol. This study was approved by the Institutional Review Board of the Chungnam National University Hospital, Daejeon, Korea (IRB No. 2019-10-053). Our IRB waived the requirement for informed consent.
Study population
We included population over 20 years of age having National Health check-ups at least one time from 2009 to 2014. We regarded the data from the first check-up as the index data, and the year of the index check-up as the index year. We included all subjects with age ³ 20 years old and excluded all subjects having pre-specified exclusion criteria. We excluded all patients previously diagnosed with HF. To assess the effect of FLI on the new-onset HF, other exclusion criteria included comorbid conditions that can affect the onset of HF including hypertension, diabetes, atrial fibrillation, cerebrovascular disease, ischemic heart disease, peripheral vascular disease, valvular heart disease, chronic kidney disease, and chronic pulmonary disease. We excluded subjects with prescribed medications including oral hypoglycemic, antihypertensive, or lipid-lowering agents within 1 year before the index check-up, subjects with increased blood pressure level above the criteria of hypertension, and elevated fasting blood glucose ³ 126mg/dL at the index check-up. We also excluded factors that can affect FLI including liver disease, and autoimmune disease before the index year. Finally, those with missing data in the index check-up were also excluded.
Definition of HF
. The primary outcome of this study was HF incidence according to FLI, and the incidence of HF was defined as the first occurrence during at least 2 different days of hospital visits, at HF admission, or death with a diagnosis of HF. We assessed each diagnosis based on the data from questionnaires, and the 1-year claim data before the index year. When we used the claims data, we defined each diagnosis as the first occurrence during at least two different days of hospital visits (outpatient) or on the first admission, as likely a diagnosis of HF. The presence of HF was defined as those with “HF” according to the ICD-10 disease code in the claim dataset. HF patients who were assigned the following ICD-10 disease codes were considered as: “hypertensive heart disease with HF” (I10.0), “hypertensive heart disease with hypertensive kidney disease with HF” (I13.0),“hypertensive heart disease with hypertensive kidney disease with HF and kidney failure” (I13.2), “ischemic cardiomyopathy” (I25.5), “dilated cardiomyopathy” (I42.0), “cardiomyopathy, unspecified” (I42.9), “cardiomyopathy in diseases classified elsewhere” (I43), “HF” (I50) including “congestive HF” (I50.0), “left ventricular failure” (I50.1), and “HF, unspecified” (I50.9).
Definition/Ascertainment of covariates
We calculated body mass index (BMI) with dividing weight (kg) by height (m)-squared. Population with a BMI of ≥ 25 kg/m2 was regarded as obesity according to the World Health Organization guideline for the Asian population.[14] Smoking status was classified into 3 categories: non-smoker, ex-smoker, and current smoker. Alcohol consumption was evaluated with using standardized self-reporting questionnaires. The questionnaires about alcohol consumption were composed of the questions asking the number of days a week alcohol is consumed, and the amount of alcohol consumed on each drinking day. The amount of alcohol consumption was calculated by multiplying them. The questionnaires about physical activity were composed of the questions asking the number of days a week 30 minutes of light exercise, 30 minutes of moderate exercise, and 20 minutes of vigorous exercise are performed, respectively. Light exercise was assumed to be 2 METs, moderate 3 METs, and vigorous 6 METs, which were multiplied by 20, 30, and 30 minutes and the respective number of days a week, and summed.
Considering the nation-wide scale of this study, laboratory test cannot be performed in a central facility. Instead, blood samples were analyzed in a number of different institutions which were qualified by an external quality assessment service, annually conducted by Korean Association of External Quality Assessment Service. Data were censored at the time of HF occurrence, disqualification of the NHIS (death or immigration), or the end of the study (December 31th, 2015).
Calculation fatty liver index
We used a well validated, surrogate marker, FLI to identify patients with NAFLD [13]. FLI was calculated with 4 variables (triglycerides [TG], BMI, gamma-glutamyl transferase [GGT], and waist circumference [WC]) with following equation:
See equation 1 in the supplementary files.
The original study showed that the FLI more than 60 as the cutoff for the diagnosis of fatty liver with positive likelihood ratio of 4.3 in general population.[15] Although the FLI is simple to calculate, and easy to screen fatty liver disease, there has been insufficient evidence regarding the diagnosis of fatty liver disease with FLI in Asians because of lower BMI and WC than other ethnic population [16]. Thus, we categorized our study group into quartiles according to their FLI and used quartile group in the statistical analysis.
Statistical analysis
We used continuous variables as mean ± standard deviation and categorical parameters as number with percentage. We performed all statistical analysis with using R software version 3.3.3 (R Foundation for Statistical Computing, Vienna, Austria; www.r-project.org). Chi-square test and one-way analysis of variance test were used to evaluate statistical differences among the FLI quartiles. The calculation of cumulative event rates according to the FLI quartiles was done with Kaplan-Meier method and compared with a log-rank test. Adjusted hazard ratios (HR) and 95% confidence interval (CI) for HF incidence were estimated with Cox proportional hazard regression analysis. In the multivariate analysis, we adjusted age and sex in the model 1, and clinical characteristics associated with new onset HF of borderline statistical significance (P<0.100) along with age and sex in the model 2. We excluded several confounding factors such as hypertension and diabetes because they have significant associations with NAFLD. The inclusion of NAFLD along with these risk factors in the multivariate model might have introduced multicollinearity into the model. We checked the multicollinearity issue by checking the variance inflation factor (VIF) in all models (Supplementary table 1). Because the VIF levels in the models were less than 10, there was no multicollinearity issue in the models. P values of < 0.05 were considered statistically significant.