Prevalence of and risk factors for fatty liver in the general population of Northern Italy: the Bagnacavallo Study

The Bagnacavallo Study was performed between October 2005 and March 2009. All citizens of Bagnacavallo (Ravenna, Emilia-Romagna, Italy) aged 30 to 60 years as of January 2005 were eligible and were invited by written letter to participate to the study. Public encounters were also held to promote participation to the study. Altered liver enzymes (ALE) were defined as alanine transaminase (ALT) > 40 U/l and/or aspartate transaminase (AST) > 37 U/l. These cut-points were the upper normal limits of ALT and AST applied by the laboratory that performed all the analyses of the Bagnacavallo study. The study protocol specified that all ALE+ and at least 50% of ALE- citizens had to undergo liver ultrasonography (LUS). ALE- citizens were chosen consecutively on the basis of their surname starting from a randomly chosen letter of the alphabet. The study was approved by the Ethical Committee of Area Vasta Romagna - IRST (reference number 112). All citizens gave written informed consent.

All participants underwent a detailed clinical history and physical examination following the model of the Dionysos Study [8, 18]. Current alcohol intake was assessed by trained interviewers by measuring the quantity (grams) of beer, wine and liquor drunk in the week prior to the enrollment [19]. Such quantity was divided by 7 to obtain a daily estimate and converted into alcohol units with rounding to the next integer. The conversion was done using an alcohol unit corresponding to 10 g of ethanol. Weight and height were measured following international guidelines [20]. Body mass index (BMI) was calculated and classified following the NIH guidelines [21]. Waist circumference (WC) was measured at the midpoint between the last rib and the iliac crest [22].

Venous blood samples were obtained after 12-h fasting. The performed blood tests included: 1) glucose; 2) triglycerides; 3) total cholesterol; 4) high-density lipoprotein (HDL) cholesterol; 5) low-density lipoprotein (LDL) cholesterol; 6) ALT; 7) AST; 8) gamma-glutamyl-transferase (GGT); 9) bilirubin; 10) hepatitis B surface antigen (HBsAg); 11) antibodies against hepatitis C virus (anti-HCV).

The MS was diagnosed using the harmonized international definition [23]. In detail, large WC was defined as WC ≥ 102 cm in men and ≥ 88 cm in women; high triglycerides as triglycerides ≥150 mg/dl or use of triglyceride-lowering drugs; low HDL as HDL < 40 mg/dl in men and < 50 mg/dl in women or use of HDL-increasing drugs; high blood pressure as systolic blood pressure ≥ 130 mmHg or diastolic blood pressure ≥ 85 mmHg or use or blood pressure-lowering drugs; high glucose as glucose ≥100 mg/dl or use of glucose lowering drugs; and MS as ≥3 of the above.

LUS was performed by five experienced physicians (ACDA, GS, FD, AL and FC) using the same methodology of the Dionysos Nutrition & Liver Study [9, 24]. Normal liver was defined as the absence of liver steatosis or other liver abnormalities. Light FL was defined as the presence of slight “bright liver” or hepatorenal echo contrast without intrahepatic vessels blurring and no deep attenuation; moderate FL as the presence of mild “bright liver” or hepatorenal echo contrast without intrahepatic vessel blurring and with deep attenuation; and severe FL as diffusely severe “bright liver” or hepatorenal echo contrast, with intrahepatic vessels blurring (no visible borders) and deep attenuation without visibility of the diaphragm. NAFLD was defined as FL associated with ethanol intake ≤2 alcohol units (20 g) / day in women and ≤ 3 alcohol units (30 g) / day in men testing negative for HBsAg and anti-HCV and not treated with steatogenic drugs [2]. AFLD was defined as FL associated with ethanol intake ≥2 alcohol units / day in women and ≥ 3 alcohol units / day in men testing negative for HBsAg and anti-HCV and not treated with steatogenic drugs [2].

Most continuous variables were not Gaussian-distributed and all are reported as median and interquartile range. Discrete variables are reported as the number and proportion of subjects with the characteristic of interest. Between-group comparisons of discrete variables were performed using Pearson’s Chi-square test and those of continuous variables using median regression with heteroskedasticity-robust standard errors [25].

Binary logistic regression was used to evaluate the association between FL and potential risk factors by means of six pre-specified models [26, 27]. The outcome of all the logistic regression models was FL (discrete; 0 = no; 1 = yes). Model 1 had ALE (0 = no; 1 = yes) as predictor; Model 2 added sex (0 = female, 1 = male) and age (continuous, years/10) to Model 1; Model 3 added BMI (continuous, kg/m²/5) and ethanol intake (continuous, alcohol units) to Model 2; Model 4 replaced BMI in Model 3 with WC (continuous, cm/10); Model 5 added MS (discrete, 0 = no; 1 = yes) and ethanol intake (continuous, alcohol units) as predictors to Model 2; Model 6 replaced MS in Model 5 with its single components, i.e. large WC (discrete, 0 = no; 1 = yes), high triglycerides (discrete, 0 = no; 1 = yes), low HDL (discrete, 0 = no; 1 = yes), high blood pressure (discrete, 0 = no; 1 = yes) and high glucose (discrete, 0 = no; 1 = yes). Before fitting the logistic regression models, we used univariable and multivariable scatterplot smoothers to get an idea of the functional form and shape of the continuous predictors and found no evidence of deviation from linearity for any predictor [28]. We also checked that the multivariable logits of the continuous predictors were linear using multivariable fractional polynomials [29]. Because alcohol intake as quantified by the present study is strictly speaking an ordinal and not a continuous variable, we tested whether its multivariable relationship with FL was linear by modeling it as both continuous and discrete in the same model [30]. We found that the relationship was linear in all models (data not shown). We also tested whether age and gender interacted in the models involving them (Models 2 to 6) and found that they did not (data not shown). Even if ethanol was not associated with FL, we nonetheless evaluated its interaction with BMI and WC because of its potential clinical significance [3]. We found no evidence of interaction of ethanol with both BMI and WC (data not shown). We evaluated the presence of collinearity among predictors in all models using the Belsley-Kuh-Welsch condition number [31]. We compared models using Akaike information criterion (AIC) and the Bayesian information criterion (BIC) and additionally calculated Nagelkerke pseudo-R² and the area the under the receiver-operating characteristic curve (-ROC-AUC) [26]. We used Model 3 to calculate and plot the sex-specific marginal probabilities of FL corresponding to the 5th, 25th, 50th, 75th and 95th internal percentiles of age and BMI at the median intake of ethanol [27, 32].

Multinomial logistic regression was used to evaluate the association between FL type and potential predictors by means of six pre-specified models [26, 27]. The outcome of all the multinomial logistic regression models was FL type (discrete; 0 = NAFLD; 1 = normal liver; 2 = AFLD). The prediction models were the same described above under binary logistic regression with the exception that ethanol was not entered into any model. We set NAFLD as the reference category in order to obtain estimates of effect sizes for the normal liver vs. NAFLD and the AFLD vs. AFLD comparisons. We performed the same tests of model assumptions described above under binary logistic regression. We compared models using Akaike information criterion (AIC) and the Bayesian information criterion (BIC) and additionally calculated Nagelkerke pseudo-R². ROC-AUC were calculated for the two binary logistic models underlying the multinomial logistic model. Statistical analysis was performed using Stata 15.1 (Stata Corporation, College Station, TX, USA).

The flow of the citizens through the study is depicted in Fig. 1. Four thousand and thirty-three (58%) of the 6920 citizens aged 30 to 60 years who resided in Bagnacavallo as of January 2005 agreed to participate to the study and 3933 (98%) of them had all the data required for analysis. The citizens were consecutively studied during the first three days of every week (except for holidays) between October 2005 and March 2009. The study protocol required that all ALE+ and at least 50% of ALE- citizens were recalled to undergo LUS.

Three hundred ninety-three (10%) of the 3933 citizens with complete data were ALE+. Sixteen (4.1%) of them had HBV or HCV infection and will not be considered here. Of the remaining 377 ALE+ citizens, 349 (93%) underwent LUS. Additional file 1: Table S1 compares the 349 ALE+ citizens with LUS to the 28 ALE+ citizens without LUS. As compared to ALE+ citizens without LUS, ALE+ citizens with LUS had higher values of triglycerides, ALT and GGT (p < 0.05). Besides not being of great interest in itself [33], the lack of statistical significance should be taken with an additional degree of caution here owing to the low number of ALE+ subjects without LUS (n = 28).

Among the 3540 (90%) ALE- citizens, 38 had HBV or HCV infection (1.1%) and will not be considered here. Of the remaining 3502 ALE- citizens, 1810 (52%) underwent LUS. Additional file 2: Table S2 compares the 1810 ALE- citizens with LUS to the 1692 ALE- citizens without LUS. As compared to ALE- citizens without LUS, ALE- citizens with LUS were more likely to be male and had higher values of age, weight, BMI, WC, glucose, triglycerides, cholesterol, LDL-cholesterol, systolic blood pressure, ALT, GGT and bilirubin (p ≤ 0.05). Besides not being of great interest in itself [33], the presence of statistical significance should be taken with an additional degree of caution here owing to the high number of ALE- subjects with (n = 1810) and without LUS (n = 1692).

Table 1 compares the ALE+ and ALE- citizens with availability of LUS. ALE+ citizens were more frequently male and slightly younger than ALE- citizens. ALE+ citizens had greater values of BMI, WC, glucose, triglycerides, total cholesterol, LDL-cholesterol, systolic blood pressure and diastolic blood pressure and lower values of HDL-cholesterol.

Table 2 reports the logistic regression models used to investigate the association between FL and potential risk factors.

Model 1 shows that the odds of FL was 5.3 (95%CI 4.1 to 6.9) times higher in ALE+ than in ALE- citizens. The corresponding probabilities of FL estimated from the logistic regression model are 74% (95%CI 70 to 79%) for ALE+ and 35% (95%CI 33 to 37%) for ALE- citizens.

Model 2 evaluates whether sex and age are associated with FL independently from ALE. While both sex (OR = 2.1, 95%CI 1.7 to 2.5 for males) and age (OR = 1.8, 95%CI 1.6 to 2.0 per decade) show an independent effect on FL, the OR of ALE changed only slightly (4%) compared to Model 1.

Model 3, obtained by adding BMI and alcohol intake as predictors to Model 2, shows that BMI is associated with FL (OR = 3.9, 95%CI 3.3 to 4.5 per 5 kg/m²) with modest changes of the OR of sex (5%) and age (11%) and with a moderate change of the OR of ALE (23%). Importantly, this model shows that ethanol intake is not associated with FL (OR = 1.0, 95%CI 1.0 to 1.1 per alcohol unit). All the employed metrics of model fit identified Model 3 as the best of all models (lowest AIC, lowest BIC, highest ROC-AUC and highest pseudo-R²).

Model 4, obtained by replacing BMI in Model 3 with WC, shows that WC is associated with FL (OR = 2.5, 95%CI 2.3 to 2.8) independently of ALE, sex and alcohol intake. The effect sizes of ALE, sex and age are similar to those of Model 3 using BMI as predictor and ethanol intake is again not associated with FL. (We did not evaluate BMI and WC together in Model 4 because of the evidence of collinearity as determined by a Belsley-Kuh-Welsch condition number of 31).

Model 5 evaluates the association of MS with FL taking into account ALE, sex and age. Having MS is associated with an odds of FL equal to 5.1 (95%CI 4.1 to 6.3). (Neither BMI nor WC were entered into this model because WC is already included in the definition of MS and BMI and WC were collinear as explained above.)

Model 6 evaluates the independent contribution of each component of the MS (large WC, high triglycerides, low HDL, high blood pressure and high glucose) to FL. Not only was each component of the MS associated with FL, but all the measures of model fit were better for Model 6 than for Model 5 (lower AIC, lower BIC, higher ROC-AUC and higher pseudo-R²), showing that there is some advantage in considering the single components of the MS instead of the MS as its association with FL is concerned.

Figure 2 plots the prevalence of FL in men and women with and without ALE as estimated by Model 3 with ethanol intake set at the median value [32]. In both sexes, the prevalence of FL increases with age and BMI.

The prevalence of NAFLD and AFLD in ALE+ and ALE- subjects was 0.46 (0.41 to 0.51) vs. 0.22 (95%CI 0.21 to 0.24) and 0.28 (0.24 to 0.33) vs. 0.13 (CI 0.11 to 0.14).

Table 3 reports the multinomial logistic regression models used to investigate the association between FL type and potential risk factors.

Not unexpectedly, all potential risk factors were associated with a lower odds of normal liver vs. NAFLD. All odds ratios were in fact < 1 for all predictors (Models 1–6). More interestingly, the same predictors were unable to discriminate AFLD from NAFLD as made clear by the unsatisfactory ROC-curves of the binary ALFD vs. NAFLD model.

In detail, ALE (Models 1–6), BMI (Model 3), WC (Model 4) and MS (Model 5) were not associated with the odds of having AFLD vs. NAFLD. Although male gender and lower age were associated with a greater odds of AFLD vs. NAFLD in all models (Models 1–6), their 95%CI are wide. It is of some interest that high WC and low HDL made AFLD less likely than NAFLD and that high triglycerides made AFLD more likely than NAFLD (Model 6) but the 95%CI of these estimates are again wide.

Although the present study focused on FL as a whole considering ethanol intake as a potential predictor, we performed an analysis of the prevalence of and the risk factors for NAFLD to allow a comparison with the available studies [10].

In the present study, the prevalence of NAFLD was 46% among ALE+ and 22% among ALE- citizens. The corresponding figures for citizens with and without suspected liver disease in the Dionysos Nutrition & Liver Study were 25 and 20% [8]. The Bagnacavallo Study and Dionysos Nutrition & Liver Study estimates are unfortunately not comparable not only because of the different operational definitions of ALE and suspected liver disease as discussed above, but also because the Dionysos Nutrition & Liver Study employed a different cut-point of alcohol intake to diagnose NAFLD in men (≤ 20 g/day) and used a different instrument (7-day prospective diary) to measure ethanol intake [8]. We have, indeed, previously shown that small errors in the estimation of ethanol intake may lead to a substantial difference in the estimated prevalence of NAFLD vs. AFLD [5, 8].

The analysis of the potential risk factors for NAFLD yielded a very interesting finding. All potential risk factors were associated with a lower odds of normal liver vs. NAFLD in all models, which were able to satisfactorily discriminate NAFLD from normal liver. However, the same models were not able to discriminate AFLD from NAFLD. There are several, not mutually exclusive, explanations for this finding. First, the dichotomization of ethanol intake, central to the separation of NAFLD from AFLD [38], could have introduced substantial bias into the inference [37]. Second, we have not cross-validated the recall method used to assess ethanol intake in the present study against an accepted reference method, e.g. the 7-day prospective diary used the Dionysos Nutrition & Liver Study [8]. Thus, we ignore the measurement error of the method used to assess ethanol intake [39]. This is unfortunately the rule in the literature on FL and is especially troublesome because the separation of NAFLD from AFLD is based entirely on the dichotomization of alcohol intake [38]. Third, it is possible that the separation between NAFLD and AFLD is not relevant at the population level because most of the risk factors for FL as whole are not associated with its separation into NAFLD and AFLD. However, without accurate measurements of ethanol intake, this hypothesis remains highly speculative. In view of the fact that ethanol is a well-known hepatotoxic agent, our data should not be taken as evidence that ethanol intake is not an individual risk factor for fatty liver but simply that with conventional instruments used to detect fatty liver and measure ethanol intake, this relationship was not evident at the population level in Bagnacavallo in 2005/2009.