Study design and population
The Fenland Study is an ongoing population-based cohort of adults from general practice lists in Cambridgeshire (Cambridge, Ely, Wisbech, and surrounding villages) in the UK [21]. Overall, 12,435 adults born between 1950 and 1975 (aged 30 to 65 years at recruitment) attended baseline clinical assessments in 2005–2015. This cohort was established to investigate environmental and genetic risk factors for the development of type 2 diabetes and related metabolic disorders. Exclusion criteria were the presence of known diabetes, being pregnant, being unable to walk unaided, or having psychosis or a terminal illness. The Cambridge Local Research Ethics Committee approved the study (04/Q0108/19), and all participants provided written informed consent.
The CoLaus Study is an ongoing population-based cohort including a random sample of 6733 individuals aged 35 to 75 years in the city of Lausanne, Switzerland. Details of the study have been described previously [22]. This cohort was established in 2003 to investigate the clinical, biological, and genetic determinants of cardiovascular disease, and it included participants of European origin. For the present analysis of CoLaus, we used data from the first follow-up (study period 2009–2013; n = 5064) when dietary assessment was initiated. The Institutional Ethics Committee of the University of Lausanne approved the study (reference 16/03), and all participants provided written informed consent.
In this analysis of the two cohorts, we used standardised exclusion criteria to remove participants with diabetes [defined as glycated haemoglobin ≥ 48 mmol/mol, fasting plasma glucose ≥ 7.0 mmol/L, 2-h glucose ≥ 11.1 mmol/L, or use of glucose-lowering drugs], those who are pregnant, and those missing dietary data, outcome data, or key covariates. We excluded those with implausible energy intake based on sex-specific thresholds (< 500 or > 3500 kcal/day in women; and < 800 or > 4000 kcal/day in men) [23]. Participants with missing marital status data (n = 1506) in Fenland were retained for analyses and coded with a missing indicator variable.
Dietary assessment
In Fenland, a self-administered, 130-item, semi-quantitative food frequency questionnaire (FFQ) was used to assess habitual dietary intake over the previous year [24]. The validity of the FFQ was previously assessed against 16-day weighed dietary record, 24-h recall, and biomarkers [25,26,27]. In this FFQ, the average consumption frequencies of each food item ranging from ‘never or less than once per month’ to ‘six or more times per day’ (nine categories) were provided. Total energy and macronutrient intake were estimated based on the UK food composition database [28].
In CoLaus, a self-administered, 97-item, semi-quantitative FFQ was used to assess dietary intake over the preceding 4 weeks [29], the validity of which had been assessed in nearby Canton of Geneva against 24-h recalls [30, 31]. For each item, consumption frequencies ranging from ‘less than once during the last 4 weeks’ to ‘two or more times per day’ (seven categories) were provided, in addition to the serving size (smaller, equal, or bigger) in comparison to a reference size.
Mediterranean diet scores
We calculated three Mediterranean diet scores (MDSs). We evaluated the pyramid-based MDS (pyrMDS) [32] based on the Mediterranean dietary pyramid [33] as our primary exposure. The Mediterranean dietary pyramid was proposed by the Mediterranean Diet Foundation to be applied to both Mediterranean and non-Mediterranean countries [33]. We previously confirmed the content validity of pyrMDS in a non-Mediterranean setting, with its higher scores being associated with lower incidence of cardiovascular disease and cardiovascular and all-cause mortality in a UK population [32]. The detailed scoring method has been reported previously [32]. Briefly, a continuous score of 0 to 1 was assigned for each recommended consumption level of the 15 components of the pyramid (possible range 0–15): vegetables, legumes, and fish as healthy food groups; red meat, processed meat, potato, and sweets as unhealthy food groups; and fruits, cereals, nuts, eggs, dairy, white meat, and alcoholic beverages as items for which a moderate consumption was recommended (Additional file 1: Table S1). The second MDS was based on an algorithm proposed by Sofi and Casini [34] based on a systematic review of the published literature (literature-based MDS, LitMDS; possible range 0–18). The LitMDS accounted for 9 dietary items: vegetables, legumes, fruits and nuts, cereals, dairy, fish, meat, alcohol, and olive oil. The third MDS was based on each cohort’s sex-specific tertiles (tMDS; possible range 0–18) and accounted for the same 9 dietary items as LitMDS (Additional file 1: Table S1). We also previously tested these scores in a UK population [32]. The MDS calculation was adjusted to an energy intake of 2000 kcal/day (8.37 MJ/day) based on the residual method [23, 32].
Ascertainment of hepatic steatosis
In Fenland, hepatic steatosis was ascertained by abdominal ultrasound, which is considered the first-line diagnostic procedure for hepatic steatosis [3]. A semi-quantitative grading system was used to define normal hepatic echotexture or mild, moderate, and severe steatosis. The images were scored retrospectively according to standardised criteria by two trained operators who were unaware of other study measures. Hepatic steatosis scoring criteria were (a) increased echotexture of the liver parenchyma (bright liver in comparison with the kidney), (b) decreased visualisation of the intra-hepatic vasculature, and (c) attenuation of the ultrasound beam. Each criterion was scored on a 4-point scale, and a cumulative liver fat score based on the sum of the scores was created (possible range 3–12) [35]. A score of ≤ 4 was classified as normal liver, 5–7 as mild steatosis, 8–10 as moderate steatosis, and ≥ 11 as severe steatosis. In our primary analysis, the mild, moderate, and severe categories were grouped together. The diagnostic accuracy of ultrasound was previously assessed against proton magnetic resonance spectroscopy, with sensitivity and specificity of 96% and 94%, respectively [36].
In both cohorts, we also evaluated fatty liver index (FLI) as an outcome, using anthropometry measures and fasting blood markers, calculated based on a logistic function including body mass index (BMI), waist circumference (WC), triglyceride (TG), and gamma-glutamyl transferase (GGT) levels as follows:
$$ \mathrm{FLI}=1/\left(1+{e}^{-\left(0.953\times \ln\ \left(\mathrm{TG}\right)+0.139\times \mathrm{BMI}+0.718\times \ln\ \left(\mathrm{GGT}\right)+0.053\times \mathrm{WC}-15.745\right)}\right) $$
FLI × 100 ranged from 0 to 100, and the presence of hepatic steatosis was defined by FLI ≥ 60, a value with sensitivity and specificity of 61% and 86%, respectively [37]. The diagnostic accuracy of FLI in comparison to ultrasonography has been reported to have an area under the receiver operating characteristic curve of 0.813 (95% CI 0.797, 0.830) [38].
In the CoLaus Study, abdominal ultrasound measurements were not available and hepatic steatosis was assessed by FLI (as above) and additionally by the ‘NAFLD liver fat score’ [39]. The NAFLD liver fat score was based on an algorithm including the presence of metabolic syndrome defined by the International Diabetes Federation criteria [40], presence of type 2 diabetes, and fasting concentrations of insulin, aspartate-aminotransferase (AST), and the AST/alanine transaminase (ALT) ratio:
$$ \mathrm{NAFLD}\ \mathrm{liver}\ \mathrm{fat}\ \mathrm{score}=-2.89+1.18\times \mathrm{metabolic}\ \mathrm{syndrome}\ \left(\mathrm{yes}/\mathrm{no}\right)+0.45\times \mathrm{type}\ 2\ \mathrm{diabetes}\ \left(\mathrm{yes}/\mathrm{no}\right)+0.15\times \mathrm{fasting}\ \mathrm{insulin}\ \left(\mathrm{mU}/\mathrm{L}\right)+0.04\times \mathrm{fasting}\ \mathrm{AST}\ \left(\mathrm{U}/\mathrm{L}\right)-0.94\times \mathrm{AST}/\mathrm{ALT} $$
Compared to proton magnetic resonance imaging, the presence of hepatic steatosis defined by a NAFLD liver fat score greater than or equal to − 0.640 had a sensitivity of 86% and a specificity of 71% [39]. In Fenland, AST levels were not available to calculate NAFLD liver fat score.
Assessment of covariates
In both studies, socio-demographic, lifestyle, and health characteristics were collected by self-administered questionnaires. Socio-demographic data included age, sex, marital status (single, married/cohabiting, and widowed/separated/divorced), occupational social class (routine/manual and administrative/professional in Fenland; working and not working in CoLaus), and educational level (compulsory, secondary, and university). Health characteristics included the presence of metabolic syndrome and family history of diabetes. In the Fenland Study, test site (Cambridge, Ely, and Wisbech) and household income (< £20,000, £20,000–40,000, and > £40,000) were also used as covariates. Smoking status was classified as ‘never’, ‘former’, and ‘current’.
In Fenland, physical activity was assessed objectively using combined heart rate and movement sensing for over 6 days (Actiheart, CamNTech, Cambridge, UK) with individual calibration for heart rate using a treadmill test [41]. To estimate intensity time series, free-living data were pre-processed and modelled using a branched equation framework then summarised over time as daily physical activity energy expenditure (kcal/day). In CoLaus, physical activity was assessed with a validated self-administered quantitative physical activity frequency questionnaire [41, 42].
In both cohorts, body weight and height were measured with participants barefoot and in light indoor clothes; BMI was calculated as weight (kg) divided by height squared (m2). Waist circumference was measured by tape mid-way between the lowest rib and the iliac crest. In Fenland, body fat mass was also measured with dual-energy X-ray absorptiometry. Blood pressure was measured three times using an automated oscillometric sphygmomanometer in both cohorts, and the average of the two last measurements was used to define systolic and diastolic blood pressure.
Fasting venous blood samples were collected. In Fenland, blood samples were placed on ice, centrifuged, and stored at − 70 °C until analysis. In CoLaus, all assays were performed on the blood samples within 2 h of blood collection. In both cohorts, plasma TG, high-density lipoprotein cholesterol, and glucose were measured using standard enzymatic methods and ALT, AST (only in CoLaus), and GGT were measured using reference method as standardised by the International Federation of Clinical Chemistry.
In both cohorts, alcohol consumption was assessed by self-report number of alcoholic beverage units consumed in the preceding week, categorised as ‘abstainers’ (0 unit/week), ‘moderate’ (1–21 units/week for men, 1–14 for women), and ‘heavy’ (> 21 units/week for men, > 14 for women) drinkers. We undertook two approaches regarding alcohol consumption: one as a covariate and the second as a component of the MDS to evaluate adherence to the Mediterranean diet from which alcohol consumption was separated out as it is an established risk factor for hepatic steatosis [1].
Statistical analysis
Statistical analyses were performed using Stata (version 14; StataCorp, College Station, TX). Descriptive statistics were presented as mean ± standard deviation (SD) for continuous variables and proportions for categorical variables. Cohen’s kappa statistics were calculated to assess the agreement between the FLI and ultrasound liver fat score in Fenland and between the FLI and NAFLD liver fat score in CoLaus.
Each of the three MDSs was modelled both categorically (quintiles) and continuously (per SD). Multivariable-adjusted Poisson regression was used to estimate the prevalence ratios (PR) and 95% confidence intervals (CI) [43] and to examine the association between hepatic steatosis (ultrasound liver fat score and FLI in Fenland; FLI and NAFLD liver fat score in CoLaus) as dependent variables and the different MDSs (pyrMDS, litMDS, and tMDS) as independent variables.
Analyses were adjusted for potential confounders including age, sex, marital status, occupational status, educational level, smoking status, energy intake, and physical activity energy expenditure (Fenland) or estimated total energy expenditure (CoLaus). Further adjustments for BMI as potential confounder were also conducted. In Fenland, the FFQ aimed at assessing habitual dietary intake across the previous year, whereas in CoLaus, it was over the past 4 weeks. Hence, to adjust for possible seasonal variation in CoLaus, the dates of dietary intake assessment were included in regression models.
A priori, we examined whether the association between adherence to the pyrMDS and hepatic steatosis varied by alcohol consumption, testing for statistical interaction by alcohol consumption and adherence to the Mediterranean diet, and we also conducted analysis stratified by alcohol consumption (abstainers, moderate, and heavy drinkers). We also assessed the influence of adjustment for both BMI and waist circumference; for body fat mass (Fenland only); for alcohol consumption (units/week); for clinical variables, including blood pressure > 130/85 mmHg (yes/no), TG level > 1.7 mmol/L (yes/no), high-density lipoprotein level < 1.29 mmol/L for men and < 1.03 mmol/L for women (yes/no), and glucose level ≥ 5.6 mmol/L (yes/no); and for family history of diabetes and metabolic syndrome (except for NAFLD liver fat score as metabolic syndrome is one of its components).
We conducted sensitivity analyses excluding the alcohol component from the pyrMDS to rule out the possible impact of alcohol on the observed association; excluding participants with BMI ≥ 30 kg/m2; excluding participants with excessive alcohol consumption; including participants with probable implausible energy intake; excluding participants with probable secondary causes of hepatic steatosis such as hepatitis B or C, HIV, or hepatotoxic medications; or including participants with diabetes (only for NAFLD liver fat score in CoLaus). The British National Formulary codes were used to identify hepatitis B and C or HIV in the Fenland cohort and the Anatomical Therapeutic Chemical classification of the World Health Organization to identify hepatotoxic medications in the CoLaus Study. Finally, we assessed the robustness of the results when modelling ultrasound liver fat score (1) using 7 as the cut-point to define hepatic steatosis (normal/mild vs. moderate/severe), and (2) continuously, we also examined the association of adherence to the Mediterranean diet with ALT and GGT levels as crude markers of hepatic steatosis [1, 44]. The latter were natural log transformed prior to regression analysis.
Possible interactions between the different MDSs with age, sex, and BMI in the main model were examined using the Wald test. We prespecified stratified analyses if significant interaction was identified. Statistical significance was considered for a two-sided test with a p value < 0.05.
