The present study is based on data from clinical studies conducted at the Sahlgrenska University Hospital, Gothenburg, Sweden. Participants (baseline characteristics presented in Table 1) were recruited to these studies from two community-based cohorts: (1) the Impaired Glucose Tolerance and Gut Microbiota Study (IGT-Microbiota study), which is a prospective observational study of Swedish people aged between 50 and 64 years. In the IGT-Microbiota study, more than 5000 men and women born in Sweden were screened to obtain a cohort of 1965 subjects with a range of glucose tolerance as assessed by an oral glucose tolerance test (OGTT)10, (2) The Gothenburg SCAPIS OGTT study, which consists of 3346 participants in the Gothenburg part of the Swedish CArdioPulmonary bioImage Study (SCAPIS) who underwent OGTT in addition to the SCAPIS protocol. SCAPIS is a prospective observational study with 30,154 individuals enrolled at ages between 50 and 64 years from a random sampling of the general Swedish population11.
In the two cohorts, imaging data on liver fat and coronary artery calcification were available in 1944 subjects in the IGT-Microbiota study and 2631 subjects in the Gothenburg SCAPIS OGTT cohort, leaving a total of 4575 subjects. Of this cohort, 322 subjects were excluded due to lipid lowering medication and a further set of 92 subjects were excluded due to missing lipid-and/or covariate measurements, leaving a total of 4161 to be included in the current study.
The studies are approved by the Ethical Review Board of University of Gothenburg, Sweden. All participants provided written informed consent. The study protocols conform to the ethical guidelines of the 1975 Declaration of Helsinki.
Liver fat and coronary artery calcification (CAC) was assessed by computed tomography (CT) scanning using a dual-source CT scanner equipped with a Stellar Detector (Siemens, Somatom Definition Flash, Siemens Medical Solution, Forchheim, Germany). CT is a well-established method to non-invasively quantify liver fat from liver attenuation values which are inversely correlated with liver fat content12. Mean liver attenuation was determined using an automatated liver segmentation algorithm13, based on a single CT slice positioned to cover both liver lobes. NAFLD was defined as liver attenuation < 50 Hounsfield Units (HU). CAC images were obtained using electrocardiogram-gated non-contrast cardiac CT imaging at 120 kV. All non-contrast image sets were reconstructed (B35f. HeartView medium CaScore) and CAC were identified and scored using the syngo.via calcium scoring software (Volume Wizard; Siemens) to obtain a CAC score according to Agatston. LDL-cholesterol (LDL-C) was directly measured using a homogenous assay (Roche Diagnostics). TRL-cholesterol (TRL-C) was defined as total cholesterol minus HDL cholesterol minus LDL cholesterol—hence this measure reflects the cholesterol content of TRL particles. ApoB was measured by an immunoturbidity, photometric method and represents total apoB (apoB100 + apoB48) in plasma. Body weight was measured with participants in light clothing, using calibrated scales, and the body mass index (BMI) was calculated by dividing the weight in kg by the square of the height in meters. Systolic and diastolic blood pressure (SBP, DBP) was registered in supine position and after 5 min of rest. The subjects fasted overnight (for at least 8 h) before the visits.
Rationale for inclusion of lipid measurements in mediation models
Since LDLs and TRLs are the main atherogenic lipoproteins in most people, the aim of the current study was to quantify the mediatory effect of LDL- and TRL particle concentration (both added together and separate) on CACS. In the current study, we had access to the following variables: apoB, plasma TG, TRL-C, LDL-C, non-HDL-C and total cholesterol. The best proxy of LDL + TRL particle concentration is total plasma apoB (which quantifies the particle concentration of LDL + TRL + Lp(a)) and the second-best proxy is non-HDL-C (which quantifies the cholesterol content of the LDL- and the TRL fraction). The closest measure of TRL particle concentration is TRL-C (cholesterol content of TRL) and plasma TG (triglycerides in plasma are mostly constituted by TRL-TG). The closest proxy of LDL particle concentration in the current study is LDL-C.
Therefore, the main variable to assess the total mediating effect of LDL + TRL particles is plasma apoB; and the main variables used to assess the mediating effect of LDL and TRL separately are LDL-C and TRL-C/plasma TG, respectively. For completeness, results for non-HDL-C and total cholesterol is also presented in the current report.
Mediation analysis quantifies the relationship between an exposure and an outcome, via the mediation of an intermediating variable14. The analysis requires two regression models. The first model regresses the exposure on the mediator, whereas the second model regresses the exposure and the mediator on the outcome. Both models should include confounders of the association as covariates. The effect of the exposure on the outcome through the mediator can be estimated by multiplying the effect of the exposure on the mediator with the effect of the mediator on the outcome, while adjusting for the exposure variable. Mediation analysis was performed in R (version 4.0.2) using the mediation package. The analyses quantified the effect of liver attenuation (modelled as a continuous variable) on CACS (modelled as 0, 1–99, 100–399, and > 400) through either plasma TG, TRL-C, LDL-C, total plasma cholesterol, and non-HDL-cholesterol (non-HDL-C). In the regression models, age and gender were used as covariates. In addition to these covariates, four more models with different sets of covariates were used as sensitivity analyses (for further description, see also Supplementary Fig. 1). These models included BMI, systolic blood pressure or smoking status in addition to age and gender, as well as a model including all five covariates. Smoking and systolic blood pressure were chosen as these may be causal factors for CAC, and may correlate with liver fat. In order to quantify the independent effects of TRLs from LDLs, multivariable models were performed using the lavaan package in R. Two multivariable models were constructed; one model including TG and LDL-C and one model including TRL-C and LDL-C. These two multivariable models were chosen because plasma TG and TRL-C are the closest proxies for TRL particle abundance, whereas LDL-C reflects LDL particle abundance. Logistic regression analysis was performed using the function glm in R, in order to quantify the relationship between NAFLD-status and (a positive) CAC score.