This study was included in the Rotterdam Study (RS) . The Rotterdam Study is a prospective, cohort, population-based study aimed at assessing the incidence and progression of risk factors for chronic disease in middle-aged and older adults. During the period 1990-1993, all residents of Ommord, a district in the city of Rotterdam, the Netherlands, aged 55 years were invited to study. 7,983 (78% of all invitees) agreed to participate (RS-I). In 2000, the group was extended with 3,011 new participants who either became 55 years old or migrated to the research area (RS-II). Participants attend follow-up exams every 3-4 years. Outcome data on morbidity and mortality were collected continuously by linking to digital files from general practitioners in the study area. .
The Rotterdam study was approved by the Erasmus MC Medical Ethics Committee (registration number MEC 02.1015) and the Dutch Ministry of Health, Welfare and Sport (WBO Population Examination Act, license number 1071272-159521-PG). The Rotterdam study has been entered into the Dutch National Trial Register (NTR; www.trialregister.nl) and at the WHO International Clinical Trials Registration Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under the common catalog number NTR6831. All participants provided written informed consent to participate in the study and to obtain their information from the treating physicians.
Body composition was assessed by DXA during the fourth visit of the RS-I (RS-I-4) and the second visit of the RS-II (RS-II-2), between 2002 and 2006 . In all, 3,724 participants were examined by DXA. We excluded participants with prevalent atrial fibrillation at baseline (n = 256). Thus, 3,468 individuals were included in the analyzes of fat depots assessed by DXA and incident atrial fibrillation. From 2003 to 2006, all participants who visited the research center were asked to participate in a computerized tomography (CT) study aimed at visualizing vascular calcification in multivascular beds. In total, 2524 participants were examined by CT scans . Of the 2,230 participants who had adequate quality CT scans (n-Excluded = 294), 91 participants with prevalent atrial fibrillation at baseline were excluded. In all, 2139 participants were included in CT-resident fat depot analyzes. Additional analyzes of body composition score and fat distribution in association with atrial fibrillation risk included participants with available DXA and CT measurements (n = 1297). (Chart 1)
Body Composition Assessment
Body composition was assessed by DXA using ProdigyTM Total body fan beam densitometer (GE Lunar Corp, Madison, WI, USA) following manufacturer’s protocols, with scans analyzed using enCORE V13.6 software (from GE Lunar) using pre-selected regions of interest. Total fat-free mass is the sum of fat-free body mass and hyperbolic lean mass (total of lean tissue from arms and legs), total fat mass is the sum of android fat mass (localized around the waist), and gynaeid fat mass (localized) around the breasts and hips and thighs) and undefined fat mass [22, 24]. As reported, the repeatability of DXA-based measurements was excellent with a Pearson correlation coefficient of 0.991 for total fat mass and 0.994 for total fat-free mass. . The ratio of fat mass from android to gynoid was also calculated. In addition to the net mass of each fat depot, we calculated the percentage fat mass, android fat mass, and genoid fat mass also by dividing each by the total body mass.
Liver fat and epicardial fat assessment
16 slices (n = 785 or 64 chips (n = 1739) A multi-detector CT scanner (Somatom Sensation 16 or 64, Siemens, Forchheim, Germany) was used to perform a non-optimized CT scan of the heart and portal ECG. Detailed information regarding imaging parameters for scans is provided elsewhere .
Liver fat was assessed using a previously described standard procedure . In summary, we have mapped three circular regions of interest in the liver using the Philips iSite Enterprise software (Royal Philips Electronics NV 2006) . These areas were plotted throughout the imaged liver tissue (including the right and left liver lobes) and carefully selected to include only liver tissue, and no irritating tissues such as large blood vessels, cysts, or focal lesions. Next, we calculated the average Hounsfield unit (HU) value from these three measurements as an indicator of the total amount of liver fat, which is a reliable proxy for the average HU value of the whole liver. . We used a fully automatic method  To determine the amount of epicardial fat per milliliter. Briefly, this method consists of whole heart segmentation and epicardial fat volume calculation. For segmentation, we used a multi-atlas approach, in which eight manually segmented (spatially aligned) contrast-enhanced cardiac scans (atlases) were recorded with a CT scan of each participant. Next, we used this segmentation to quantify epicardial fat.
Atrial fibrillation assessment
Methods for making judgments in events related to prevalent and incident atrial fibrillation have been described previously [13, 22, 28, 29]Ascertainment of atrial fibrillation at baseline and follow-up examinations in our study was based on clinical information from the medical records of all participants in the Rotterdam study. Within the Rotterdam study, data on medical history and medication use is continuously collected through multiple sources including the baseline interview at home, physical examination at our research center, pharmacy prescription records, the national medical record of all primary and secondary hospital discharge diagnoses, and screening General Practitioner Records. In addition, a 10-s 12-lead resting electrocardiogram (ECG) used with the ACTA Gnosis IV ECG recorder (Esaote; Biomedicine, Florence Italy) is obtained from all participants at each Rotterdam Study visit to check for fibrillation atrial; ECG records were digitally stored and analyzed using the Standardized Electrocardiogram Analysis System (MEANS). ECG records were digitally stored and analyzed using the Standardized Electrocardiogram Analysis System (MEANS). Then, the AF findings are adjudicated independently by two research clinicians. In case of disagreement, a senior cardiologist is consulted. Participants were followed from the date of RS enrollment until the date of onset of atrial fibrillation, the date of death, loss to follow-up, or until January 1, 2014, whichever occurred first.
Assessment of cardiovascular risk factors
Methods for assessing cardiovascular risk factors are detailed in Supplemental File 1: Methods [22, 28, 30].
Descriptions were presented as mean ± standard deviation (SD) or median (interquartile range – IQR) for the variables and continuous numbers (percentages) for the categorical variables. As values of HU (a) had a left skewed and abnormal distribution, we used exponential transformed values (B) [B = A3.5/10,000] . Also, since a lower value of HU represents a greater amount of liver fat, we inverted the values of HU (-HU) to represent levels of liver fat.
The values for each fat depot were standardized to allow direct comparisons. Cox proportional hazards regression analysis was used to examine whether different measures of fat depot including fat mass, android fat mass, gynoid fat mass, lipid-to-genoid ratio, hepatic fat, and epicardial fat are associated with the onset of new atrial fibrillation. Besides, in the analysis with DXA metrics, we also assessed each percentage of fat mass, android fat mass, and genoid fat mass of total body mass in association with incident AF. In the first model, we calculated the age- and sex-specific hazard ratio (HR) and 95% confidence intervals (CIs). Model 2 was also adjusted for cardiovascular risk factors including total and HDL cholesterol, history of hypertension, history of diabetes (DM), history of coronary heart disease (CHD), history of heart failure (HF), history of left ventricular hypertrophy, status of Smoking, total alcohol intake, use of lipid-lowering medications, and use of heart medications. Furthermore, we adjusted the total fat-free mass in Model 3 and the total fat mass in Model 4, respectively. Finally, the potential nonlinear associations between different fat depots and incident AF were examined using slice analyzes in Cox models. However, there was no indication of significant nonlinearity (all s For nonlinearities > 0.05, the result is not shown). In sensitivity analyses, effect modification by sex was tested. Besides, analyzes were repeated after participants were stratified by BMI. Categories BMI <25 and BMI greater than 25 kg/m2. We also performed all analyzes among participants without predominant cardiovascular disease (CHD, HF, and stroke) at baseline.
To explore the potential cumulative effect of fat stores, we also included the five fat depots (fat mass, android fat mass, genoid fat mass, liver fat, and epicardial fat) to generate a body fat score. Each fat depot was scored from 0 to 2 according to the respective third. All component scores were combined to obtain a total score of 0-10 in our population, with higher scores indicating higher total body fat. Thus, Cox proportional hazards regression analyzes were performed, with the same adjusted multivariate models as above, to investigate the effect of body fat score (eg tertiles: score 0-2, score 3-6, score 7-10 with the first tertile as a group Bookmark) on the incident AF.
Then we developed fat distribution patterns. Distribution patterns were derived using Principal Component Analysis (PCA) on fat depot values (including fat mass, android fat mass, gynaeid fat mass, liver fat, and epicardial fat) . We used a varimax rotation to obtain the potential principal components. Factor loading, which reflects the standard association between each fat depot and body fat distribution pattern, was used to characterize potential patterns using a cut-off of 0.5. For each participant, pattern adherence scores were generated by summarizing the observed values of the pattern’s lipid depots weighted by loading the corresponding factor for each of the patterns separately. Furthermore, Cox proportional hazards regression analysis was used to assess how specific fat distribution patterns (eg interquartiles with the first quartile as a reference) relate to incident atrial fibrillation.
All missing values in covariates were calculated assuming none at random using multiple imputation . For multiple imputation, all available data were used to generate 5 imputed data sets. Statistical significance was considered at the tail s-value < 0.05. Analyzes were performed using R software (R 4.0.3; R Foundation for Statistical Computing, Vienna, Austria).