Objective A healthy lifestyle is the first-line treatment in non-alcoholic fatty liver disease (NAFLD), but specific dietary recommendations are lacking. Therefore, we aimed to determine whether dietary macronutrient composition is associated with NAFLD.
Design Participants from the Rotterdam Study were assessed on (1) average intake of macronutrients (protein, carbohydrate, fat, fibre) using a Food Frequency Questionnaire and (2) NAFLD presence using ultrasonography, in absence of excessive alcohol, steatogenic drugs and viral hepatitis. Macronutrients were analysed using the nutrient density method and ranked (Q1–Q4). Logistic regression analyses were adjusted for sociodemographic, lifestyle and metabolic covariates. Moreover, analyses were adjusted for and stratified by body mass index (BMI) (25 kg/m2). Also, substitution models were built.
Results In total, 3882 participants were included (age 70±9, 58% female). NAFLD was present in 1337 (34%) participants of whom 132 were lean and 1205 overweight. Total protein was associated with overweight NAFLD after adjustment for sociodemographic and lifestyle covariates (ORQ4vsQ1 1.40; 95% CI 1.11 to 1.77). This association was driven by animal protein (ORQ4vsQ1 1.54; 95% CI 1.20 to 1.98). After adjustment for metabolic covariates, only animal protein remained associated with overweight NAFLD (ORQ4vsQ1 1.36; 95% CI 1.05 to 1.77). Monosaccharides and disaccharides were associated with lower overall NAFLD prevalence (ORQ4vsQ1 0.66; 95% CI 0.52 to 0.83) but this effect diminished after adjustment for metabolic covariates and BMI. No consistent associations were observed for fat subtypes or fibre. There were no substitution effects.
Conclusion This large population-based study shows that high animal protein intake is associated with NAFLD in overweight, predominantly aged Caucasians, independently of well-known risk factors. Contrary to previous literature, our results do not support a harmful association of monosaccharides and disaccharides with NAFLD.
- fatty liver
Statistics from Altmetric.com
Significance of this study
What is already known on this subject?
Lifestyle modification is the cornerstone of treatment in non-alcoholic fatty liver disease (NAFLD). Beneficial effects from weight loss of 5% or more have been repeatedly observed. However, specific evidence-based dietary recommendations on dietary composition for NAFLD are lacking.
Current evidence on dietary intake in relation to NAFLD originate from small studies of selected patient groups with often suboptimal methodology and a main focus on dietary fats and carbohydrates. Furthermore, although the mono-unsaturated fatty acid-rich Mediterranean Diet has been advocated for NAFLD, the evidence for this diet is limited.
What are the new findings?
In this well-defined prospective cohort, 3882 elderly Caucasian participants completed extensive Food Frequency Questionnaires and underwent ultrasound evaluation to diagnose presence of NAFLD.
We found animal protein intake to be associated with greater prevalence of NAFLD in overweight participants independent of energy intake and numerous other risk factors.
Also, our results did not indicate a harmful association of monosaccharides and disaccharides with NAFLD, although within the context of this study population with limited sugar consumption.
How might it impact on clinical practice in the foreseeable future?
The results of this large study add to the current evidence on the importance of dietary composition in NAFLD independent of caloric intake.
In particular, it shifts focus from the carbohydrate and fat debate towards the third, previously underexplored macronutrient, protein.
Since the first case description by Ludwig et al in 1980, the prevalence of non-alcoholic fatty liver disease (NAFLD) increased expeditiously paralleling the obesity epidemic.1 2 NAFLD is characterised by fat deposition in the liver in absence of excessive alcohol consumption or established liver disease.2 It is referred to as the hepatic manifestation of the metabolic syndrome and has now become the most common liver disease affecting an estimated one-third of adults in the general population of developed countries.3 In high-risk populations with type 2 diabetes and metabolic comorbidities, prevalence of NAFLD even reaches up to 70%.4 Progression of NAFLD can lead to fibrosis, cirrhosis, hepatocellular carcinoma and liver failure with corresponding life-threatening complications. The end-stage disease often requires liver transplantation. Indeed NAFLD already constitutes the second most common indication for liver transplantation in the USA and is predicted to become the number one indication soon.5 In addition to the above-mentioned, NAFLD also contributes to an increased risk for metabolic and cardiovascular morbidity and mortality.6 Hence, NAFLD has emerged as a great global health threat and subsequently, prevention and treatment thereof are of strong public interest.
NAFLD is more common in people with an unhealthy lifestyle, that is, with an unhealthy diet and physical inactivity.2 Although there are several hundreds of promising pharmacological trials ongoing, there is still no registered drug for the treatment of NAFLD. Therefore, in daily practice, lifestyle modification remains the first-line treatment in NAFLD.7 At present, weight loss of 5%–7% or more is recommended, based on two prospective trials in overweight patients.8 9 However, not all overweight individuals will develop NAFLD, and likewise, not all individuals with NAFLD are overweight.10 This gives food for thought on whether dietary quality, rather than dietary quantity, is important in the pathogenesis and treatment of NAFLD. Current dietary recommendations include caloric restriction and adherence to the macronutrient composition of the Mediterranean diet.7 However, evidence on the mono-unsaturated fatty acid (MUFA)–rich Mediterranean diet for NAFLD is limited by small study populations (n=12–90 subjects), suboptimal nutritional analyses or use of surrogate primary endpoints (ie, liver transaminases) rather than imaging diagnosis of NAFLD.11 Moreover, health recommendations on fat and carbohydrate consumption have been widely debated.12 13 Only a minority of studies examined the effect of all macronutrients combined, and those who did, showed conflicting results.14–17 In addition, these studies too are hampered by small sample size (n=56–349 subjects) and/or by suboptimal methodology (eg, not correcting for energy intake, body mass index (BMI), overall dietary quality or other potential confounders).
So far, no study has examined macronutrient composition in relation to NAFLD on a large scale using comprehensive nutritional analyses methods including energy density and substitution models, taking into account potentially important sociodemographic, lifestyle and metabolic risk factors. We therefore conducted a large population-based study in elderly Caucasians, who completed a validated 389-item Food Frequency Questionnaire (FFQ) and underwent hepatic ultrasound, to determine whether macronutrient intake is associated with NAFLD independently of total energy intake and a large number of potentially confounding traits.
The Rotterdam Study (RS) is a large ongoing population-based cohort of predominantly elder participants residing in a suburb of Rotterdam, the Netherlands. The design and rationale of this population-based study have been described in detail previously.18 In short, the study commenced in 1989 and comprises three different cohorts (RS I, RS II and RS III). All residents aged 55 (RS I, RS II) or 45 (RS III) and above were invited to participate. Participation rate of these cohorts were 78%, 67% and 65%, respectively. Liver imaging is part of the core protocol since 2009. Hence, all participants who underwent abdominal ultrasound between January 2009 and June 2014 were included. Written informed consent was obtained from the participants.
Participants were asked to complete an externally validated semiquantitative 389-item FFQ developed for Dutch adults during their visit at the research centre.19 20 This questionnaire included detailed questions on food item consumption over the last month and addressed frequency, portion size, type of food item and preparation methods. Servings were estimated in grams per day using standardised household measures,21 and macronutrient intake was extracted from the questionnaires using the Dutch Food Composition Table (NEVO v2011) that includes information on nutrient content per gram or serving per product. Incomplete or unreliable FFQs, defined as caloric intake of less than 500 or more than 7500 kilocalories (kcal), were excluded. To correct for potential measurement error and to examine the relative contribution of a macronutrient to the diet, we adjusted for energy intake using the nutrient density method.22 For example, 1 g of protein equals 4 kcal, hence to calculate the energy per cent of protein (E%) = (total protein intake (g)×4/total kcal intake)×100. Similarly, the E% of carbohydrates (4 kcal/g), fats (9 kcal/g), fibre (2 kcal/g) and alcohol (6 kcal/g) were determined. Subsequently, all E% were ranked into quartiles (Q1=lowest quartile).
Additionally, to account for confounding by overall dietary quality, the Dutch Healthy Diet Index (DHDI) was derived from the FFQ and added to the multivariable analyses.23 A higher DHDI indicates stricter adherence to the Dutch dietary guidelines. DHDI is an adherence score designed specifically for the Netherlands, but correlates highly with the perhaps more familiar (Alternate)- Healthy Eating Index, the (A)HEI (r≥0.60).24 For the purpose of this study, the DHDI was modified in multivariable models to avoid multicollinearity (eg, macronutrient analyses of fat is adjusted for DHDI minus the trans fatty acid and saturated fat components of the DHDI).
Steatosis was assessed using abdominal ultrasound, which was carried out by a certified and experienced technician on Hitachi HI VISION 900. All participants were unaware of the presence of steatosis before completing the FFQs. Ultrasound images were stored digitally and re-evaluated by a single hepatologist with over 10 years of experience (RK). The ultrasound technician and hepatologist were blinded for the FFQ data. Diagnosis of steatosis was determined dichotomously according to Hamaguchi et al,25 as presence or absence of hyperechogenic liver parenchyma. Participants with possible secondary causes for steatosis were excluded from this study, that is, (1) excessive alcohol consumption (>30 g/day for men and >20 g/day for women) as assessed by the FFQ; (2) use of steatogenic drugs, that is, amiodarone, systemic corticosteroids, methotrexate or tamoxifen, extracted from linked pharmacy data and (3) viral hepatitis, based on hepatitis B surface antigen and anti-hepatitis C virus serology, as measured by an automatic immunoassay (Roche Diagnostic GmbH). Of note, for participants from RS-III, there was a median time-gap of 5.5 years between completing the FFQ (before introduction of liver ultrasound) and the performance of ultrasound. Because dietary data are known to be stable over time, RS-III was included in the total study population.26
Biochemistry and additional covariates
All blood samples were collected after overnight fasting just before abdominal ultrasound. Blood lipids, platelet count, glucose, alanine aminotransferase, aspartate aminotransferase, gamma-glutamyl transferase and total bilirubin were measured using automatic enzyme procedures (Roche Diagnostic GmbH, Mannheim, DE). Insulin was determined using an automatic immunoassay (Roche Diagnostic GmbH).
Data concerning demographics, physical activity, smoking, educational level and comorbid conditions were obtained during an extensive home interview by trained interviewers. Anthropometric measurements were carried out by well-trained research assistants measuring height (m), weight (kg) and waist circumference (WC in cm). Blood pressure measurements (mm Hg) were obtained at a single visit using two subsequent measurements in upright position after a minimum of 5 min rest. BMI was calculated as weight/height2 (kg/m2) and considered lean if <25 kg/m2 and overweight if ≥25 kg/m2.
Insulin resistance was assessed using the homeostasis model assessment of insulin resistance, as fasting glucose (mmol/dL) multiplied by fasting insulin (mU/L) divided by 22.5.27 The metabolic syndrome was diagnosed when at least three of the following traits were present: (1) abdominal obesity, defined as WC≥102 cm in men and ≥88 cm in women; (2) serum triglycerides≥150 mg/dL (1.7 mmol/L) or drug treatment for elevated triglycerides; (3) serum high-density lipoprotein cholesterol (HDL-C) ≤40 mg/dL (1.0 mmol/L) in men and ≤50 mg/dL (1.3 mmol/L) in women, or drug treatment for low HDL-C; (4) blood pressure ≥130/85 mm Hg or drug treatment for elevated blood pressure and (5) fasting plasma glucose ≥100 mg/dL (5.6 mmol/L) or drug treatment for elevated blood glucose.28
After excluding participants with either missing or unreliable FFQs and participants with more than 30% missing study variables, the remaining missing values (range of 0.02%–10.79% within covariates) were imputed using multiple imputation (fully conditioned specification) to reduce bias due to missing data.29 Ten imputed datasets were created using the R-package mice and the analyses were performed in each dataset before the results were pooled by Rubin’s rules, to take into account uncertainty with the prediction of missing data.30 The imputation process is described in more detail in the online supplementary methods.
Descriptive statistics were used to describe population characteristics. Continuous data were presented as mean±standard deviation (SD) or median with 25th and 75th percentile (P25 to P75) according to the distribution of the variable. Categorical data were presented as percentage. χ² test, Student’s t-test or Wilcoxon Rank Sum test were used to evaluate differences in categorical, normally distributed and not-normally distributed data, between subjects with and without NAFLD. In order to give more insight in the composition of the different macronutrients, we created 49 relevant food groups out of the 389 food items and performed a Spearman correlation to identify the three topmost correlations between macronutrients and food groups.
All macronutrients were analysed continuously (in E%) using standardised values (increase per 1SD) as well as in quartiles using Q1 as reference. We used three separate multivariable logistic regression models to assess the associations of macronutrients with NAFLD. The first ‘sociodemographic’ model (model 1) includes adjustment for age, gender, education level (low/moderate/high) and study cohort. The second ‘lifestyle confounding’ model (model 2) is additionally adjusted for smoking status (never vs past/current), alcohol in E%, energy intake (kcal), physical activity (metabolic equivalent of task (MET) hours/week) and DHDI. Finally, model 3, the’ metabolic’ model, was additionally adjusted for presence of diabetes, metabolic syndrome and total cholesterol (mmol/L). Moreover, analyses with carbohydrates were adjusted for fibre intake and vice versa, and all subtypes of one macronutrient were adjusted for each other. Results were presented as OR with 95% CI.
BMI is an important covariate; it could act as potential mediator (in the pathway between the exposure (diet) and the outcome (NAFLD)), as a confounder (affecting both exposure and outcome, causing a false association), as a collider (a covariate that is not in the pathway but is influenced by both the exposure and the outcome, it may create a non-existing association between the exposure and outcome) or as effect modifier (indicating different associations in subgroups of patients). Also, participants with NAFLD and a normal BMI (lean NAFLD) could have a different pathophysiological pathway from overweight NAFLD, for example, through genetic predisposition or body composition. In addition, measurement error, eating habits and, hence, macronutrient associations with NAFLD could have differed between lean and overweight individuals. Therefore, we used the following approaches to account for BMI: first, we evaluated the linearity of BMI in model 3 in relation to steatosis using cubic splines and found a non-linear effect using log-likelihood ratio testing. The figure of the spline showed that a log-shaped form could improve the fit of the model, we therefore created a model 4, adding log-transformed BMI as covariate to the metabolic model to evaluate changes in effect estimates. We believe the effect of BMI can be studied best in this model 4, because here we already adjusted for all other potential confounding factors. Second, we tested for interaction between BMI and each macronutrient (eg, BMI×total protein). And third, all analyses were stratified for lean and overweight participants at a cut-point of 25 kg/m2, while this is a widely used cut-off as established by the WHO and because there was no clear cut-off point visible in the cubic splines.
Also, to evaluate whether the observed associations were due to higher intake of that specific macronutrient rather than lower intake of another macronutrient, we performed substitution analyses in the metabolic model.31 For example, the substitution model of replacing total protein for total fat included the following dietary covariates, total protein, total carbohydrates, total fibre and alcohol, but not total fat, in addition to the above-mentioned metabolic and environmental covariates. The obtained estimate from the regression coefficient for protein from this model reflects the theoretical effect of replacing all fat intake completely with protein intake (in E%). Additional sensitivity analyses were performed comparing the analyses of imputed data to that of the complete case and assessed differences between the group with and without missing cases in order to assess the robustness of our data. In addition, we excluded the third cohort from the final analyses as they filled in their FFQ 5.5 years prior to liver imaging.
To correct for the inflated type I error that arises due to multiple testing we applied the method proposed by Sidák,32 adapted as described in Galwey et al,33 using the effective number of tests instead of the actual number of tests. This adaptation is necessary to take into account that dietary exposures intercorrelate (ie, the intake of individual macronutrients are not fully independent from each other) and, hence, the corresponding tests are not independent from each other. The resulting corrected significance level for all macronutrient analyses was p<0.021. All analyses were performed using SPSS V.21.0 (SPSS, Chicago, Illinois, USA) and R V.3.4.1.
Participants were not involved in setting the research question or the outcome measures, nor were they involved in developing plans for recruitment, design or implementation of the study. Participants were not asked for advise on the interpretation of results. All participants were regularly updated on study outcomes via a home-sent newsletter and the study website.
The flowchart of the study population is depicted in figure 1. A total of 5967 participants were eligible for this study. First, we excluded participants due to missing, incomplete or unreliable FFQs (n=1173; 19.7%). These participants were younger (mean age 68.5 vs 69.6 years old; p<0.01), less often of Caucasian origin (95.0% vs 98.0%; p<0.01) and had a slightly higher BMI (27.5 vs 26.9 kg/m2; p<0.01) than the included study group, whereas gender and frequency of steatosis were similar (55.8% vs 57.5% female; p=0.27% and 37.2% vs 35.5% steatosis; p=0.27). Second, we excluded participants with more than 30% of missing study variables (0.8%) and participants with secondary causes for steatosis (18.3%). Hence, the total study population included 3882 participants. Population characteristics are presented in table 1 and original and imputed data in online supplementary table 1 In short, mean age was 69.7±8.8, 58.3% were female, the majority of participants were of Caucasian origin (97.6%) and median BMI was 26.9 (24.5–29.7) kg/m2. NAFLD prevalence was 34.4% (n=1337). Participants with NAFLD had lower education level, were more often current or former smokers, performed less physical activity, had higher BMI, more comorbidities and more deviant mean or median laboratory values although within the normal range.
Dietary characteristics are presented as relative consumption (E%) in table 1. Animal protein correlated most with red meat (r=0.35), refined meat (r=0.25) and fish (r=0.24) and vegetable protein with whole grain, rice, bread and pasta (r=0.50), cruciferous vegetables (r=0.27) and vegetarian food products (r=0.24). Monosaccharides and disaccharides correlated most with fruit (r=0.63), sweets (r=0.30) and fruit juice (r=0.24) and polysaccharides with whole grain, rice, bread and pasta (r=0.39), refined grain, rice, bread and pasta (r=0.29) and potatoes (r=0.27). Last, saturated fat correlated most with full fat cheese/cream (r=0.50), full/non fluid fat (r=0.42) and desserts (0.27); MUFAs with sauce (r=0.29), fried snacks (r=0.29) and oil fats (r=0.29); poly-unsaturated fatty acids with diet/fluid fats (r=0.35), full/non fluid fats (r=0.32) and peanuts (r=0.25) and trans fatty acid with full fat cheese/cream (r=0.39), full/non fluid fat (r=0.38) and desserts (r=0.32). NAFLD participants reported lower median caloric consumption than participants without NAFLD (1996 vs 2052 kcal, table 1). The same was observed for BMI (overweight individuals reported 2006 kcal vs 2089 kcal in lean individuals; p<0.01). NAFLD participants reported higher median total protein, animal protein, total fat, saturated fat and trans fatty acid intake (all in E%, table 1). Moreover, they reported lower total carbohydrate and monosaccharide and disaccharide consumption and marginally lower fibre consumption (table 1). Absolute consumption of macronutrients in grams and energy percentage per quartile are given in online supplementary table 2.
Protein consumption and NAFLD
Both total and animal protein were associated with higher odds for NAFLD in the first three models (table 2). Vegetable protein was not associated with NAFLD in any of the models. After adjustment for log-transformed BMI none of the associations remained. However, effect modification by BMI was suggested: interaction with BMI was p=0.10 for total protein, p=0.04 for animal protein and p=0.19 for vegetable protein. Indeed, stratified analyses by BMI revealed that both total protein and animal protein intake were associated with overweight NAFLD in models 1 and 2 (figure 2A). Vegetable protein was associated with overweight NAFLD as well, but only in model 2. In model 3, the association with overweight NAFLD attenuated but remained significant for animal protein (ORQ4vsQ1 1.36, 95% CI 1.05 to 1.77; pfor trend=0.09) while not for total protein (ORQ4vsQ1 1.22, 95% CI 0.95 to 1.55; pfor trend=0.09), as shown in figure 2B. Detailed results of the stratified models are displayed in online supplementary table 3.
Carbohydrate consumption and NAFLD
Total carbohydrate intake and monosaccharide and disaccharide intake were inversely related with NAFLD prevalence in both models 1 and 2, but the associations attenuated in the third and fourth models (table 2). In contrast, polysaccharide consumption was not associated in the first two models, but after adjustment for metabolic traits in model 3, there was an inverse association with NAFLD (table 2). Finally, after correction for log-transformed BMI, none of the associations remained significant. Also, no association between dietary fibre and NAFLD was found. The associated did not differ by BMI (p for interaction was p=0.08 for total carbohydrate, p=0.45 for monosaccharides and disaccharides, p=0.16 for polysaccharides and p=0.55 for fibre intake). Comprehensive results of the stratified models are shown in online supplementary table 3. As depicted in figure 2A, only monosaccharides and disaccharides were significantly associated with lower odds for NAFLD in overweight individuals. Yet, direction and magnitude of estimates in lean participants were comparable to those of overweight participants (OR 0.72 for both, model 2). After adjustment for metabolic covariates in model 3, the association for monosaccharides and disaccharides in overweight dissipated (ORQ4vsQ1 0.83, 95% CI 0.63 to 1.10; pfor trend=0.28, figure 2B and online supplementary table 3).
Fat consumption and NAFLD
PUFA consumption was associated with lower odds for NAFLD in the first three models, but results diminished after adjustment for log-transformed BMI (table 2). No effect modification by BMI was observed (p for interaction was p=0.16 for total fat, p=0.20 for saturated fat, p=0.21 for MUFAs, p=0.54 for PUFAs and p=0.96 for trans fatty acids). Figure 2A,B shows the stratified models 2 and 3, in which PUFA intake was no longer associated with NAFLD (online supplementary table 3). Saturated fat became associated with lean NAFLD after metabolic adjustment (ORQ4vsQ1 2.21, 95% CI 1.03 to 4.72; pfor trend=0.03, figure 2B and online supplementary table 3).
We did not observe consistent substitution effects when one (sub)type of macronutrient was substituted for another (sub)type of macronutrient (online supplementary table 4).
First, 3259 participants had complete data on all variables, and 623 had missing data on at least one covariate. Estimates derived from the complete case analyses were more pronounced than the imputed analyses, with all results pointing in the same direction. For example, in the imputed analyses (online supplementary table 3) animal protein had an ORQ4vsQ1 of 1.36 (95% CI 1.05 to 1.77; pfor trend=0.09) for overweight NAFLD compared with an ORQ4vsQ1 of 1.52 (95% CI 1.14 to 2.03; pfor trend=0.03) in the complete case analyses (online supplementary table 5). We therefore compared the group with complete cases to the group with missing data. In the latter group (total n=623), physical activity was the variable most often missing before imputation (n=358), followed by smoking status (n=203) and education level (n=50). The participants with missing variables were older (71.1 vs 69.5) and were less often female (47% vs 60%). Moreover, they had higher prevalence of diabetes (17.3% vs 12.4%) and NAFLD (38.4% vs 33.7%) and had a higher BMI (27.4 vs 26.8 kg/m2). More detailed information is shown in online supplementary table 6.
Second, we excluded the third cohort from the final analysis in order to avoid possible bias induced by a time lag of 5.5 years (median) between completion of the FFQs and liver imaging. The direction of the results did not change, but significance attenuated as can be seen in online supplementary table 7. However, the results should be interpreted in light of decreased power and the difference in cohort characteristics, such as age (mean age RSI and II was 75.3 vs 62.0 years in RS III). Third, additional adjustment for coffee as potential confounding covariate did not change the association between macronutrients and NAFLD (data not shown).
This is the first large population-based cohort study to examine macronutrient intake using an extensive and externally validated semiquantitative FFQ in relation to ultrasound-confirmed NAFLD. The results of this cross-sectional analysis, with FFQs preceding ultrasound, imply that specific macronutrients are associated with NAFLD independent of energy intake. Specifically, high animal protein intake was associated with higher prevalence of NAFLD in overweight participants. In addition, we found a trend towards lower prevalence of NAFLD in those with high consumption of monosaccharides and disaccharides. However, this association did not hold true after adjustment for log-transformed BMI. We did not observe consistent substitution effects of macronutrient replacement, emphasising the need of a diverse diet.
Recent dietary review papers on NAFLD have advocated implementation of the Mediterranean diet, which is rich in MUFAs, fruits, legumes and nuts and low in saturated fat, carbohydrates and red meat.34 Although we analysed diet based on macronutrient composition and not on predefined dietary patterns, we found that intake of animal protein was significantly associated with overweight NAFLD independent of sociodemographic, lifestyle and metabolic traits. Furthermore, we found that the association between animal protein and NAFLD is mainly present in the highest quartile and does not appear to be dose-responsive.
Our findings are in line with previous studies, which showed that patients with NAFLD consumed significantly more meat than controls35 even after adjustment for confounders and energy intake.14 Moreover, another recent Dutch population study found similar results, showing higher intake of protein from animal sources in individuals with a fatty liver, identified by the Fatty Liver Index (a non-invasive algorithm) rather than liver imaging.17 In this study however, BMI was not taken into account as a covariate, and macronutrients were not adjusted for dietary quality (eg, DHDI). Interestingly, a large epidemiological study showed that high red meat intake was associated with all-cause mortality and in particular with mortality from liver diseases (HR 2.30, highest vs lowest quintile).36 In addition, a study from Israel found that high meat consumption, specifically high red and processed meat consumption were associated with NAFLD and insulin resistance, independent of saturated fat intake and BMI.37 Yet, two other studies did either not find a difference in protein consumption between patients and controls or found that controls consumed slightly more protein than patients.15 16 However, both studies used absolute consumption in grams instead of energy adjusted intake and did not distinguish between animal or vegetable protein. Another recent study suggested a beneficial effect of protein. This was an intervention study in 37 diabetics with mild steatosis (<30% lipid content on MRI), in whom intrahepatic lipid content reduced on a strict high vegetable or animal protein diet for 6 weeks.38 Since this study differs from ours in various ways (ie, our study included n=3882 individuals with a low prevalence of diabetes (13%) in an observational rather than an interventional study design with an outcome defined by >30% steatosis as set by the detection limits of ultrasonography), direct comparison is difficult.
Contrary to common belief, we did not find a ‘harmful’ association between carbohydrates and NAFLD. In contrast, participants with high monosaccharide and disaccharide intake initially showed lower odds for NAFLD, but this association attenuated after BMI adjustment. Although the general assumption is that fructose intake harms the liver, evidence for this assumption is indeterminate.39 In most studies, it is difficult to separate the contribution of fructose-containing sugars from that of other dietary factors, such as origin of food item, energy intake and overall dietary quality.13 This lack of clarity is supported by experimental studies showing that isocaloric carbohydrate intake was not associated with steatosis, but rather with amount of calories.40 41 Most studies to date have shown detrimental effects on NAFLD, but focused only on fructose intake from soft drinks.42 43 In fact, in this predominantly elderly population median soft drink consumption was less than 1 glass per day (ie, 44.5% consumers that have a median consumption of 0.36 glasses/day (0.14–0.91)). Indeed, the main food group contributing to monosaccharides and disaccharides in this population was actually fruit, and soda was not among the top 3 correlated contributors. This may partly explain why we did not observe a negative association.
Our results also do not suggest a ‘beneficial’ role for MUFAs. This is in line with previous studies on macronutrient associations with NAFLD, which either did not find an association with MUFAs or found higher MUFA consumption in the NAFLD-group than in controls.14–17 In contrast, a randomised controlled trial from Italy showed a reduction in liver fat following a MUFA enriched-diet in patients with diabetes.44 However, this study was small (nine participants/arm) and showed a significant but small decrease in liver fat percentage (2.2% in the MUFA-arm). In addition, MUFA-intake in this study was much higher (~25%) than in our study (10.7%).44 Dietary fat consumption, in particular saturated fat, remains a widely debated topic45 and evidence on associations of (subtypes of) fat with incident metabolic disease is heterogeneous and suffer from residual confounding.46 Our substitution analyses did imply a favourable trend when substituting PUFAs for animal and total protein intake in overweight participants. However, this association was not significant.
There are several possible mechanistic explanations as to how high animal protein intake could be associated with overweight NAFLD. Although we adjusted for overall dietary quality, the association might be explained by other dietary components. One hypothesis is that constituents, such as nitrate, nitrite, heme iron and their by-products, in both unprocessed and processed meat could act as mediators between dietary intake and cardiovascular and metabolic homeostasis.36 Heme iron is associated with increased oxidative stress and insulin resistance.47 Nitrate and nitrite have been associated with endothelial dysfunction and insulin resistance.48 Moreover, a large prospective cohort study found that nitrate, nitrite and heme iron from red meat intake were all associated with higher risk of chronic liver disease.49
Another possible mechanism through which animal protein could be associated with NAFLD is low-grade metabolic acidosis induced by a high diet-dependent acid load. The Western diet, characterised by high intake of acidic food items (eg, animal protein) and low intake of alkali, potassium-rich food items (eg, vegetables/fruits) increases daily acid load.50 Recently, diet-dependent acid load has been associated with a higher risk of NAFLD.35 51 The authors of these studies hypothesised that high dietary acid load might suppress growth hormone secretion and subsequent insulin-like growth factor-1 response, which both have been associated with NAFLD. In addition, some experimental studies have showed that high dietary acid load reduces extracellular pH and insulin sensitivity and decreases beta cell response.52 This could lead to diabetes and NAFLD, as insulin resistance is the key dysfunction in this disease.2
The main strength of this study is the use of a large unselected study population and a robust statistical analysis with correction for a great number of sociodemographic, lifestyle and metabolic traits, as assessed by well-validated tools. Furthermore, we used widely accepted nutritional epidemiological methods and performed sensitivity analyses, emphasising the robustness of our results. Also, we corrected for multiple comparisons using Sidák-corrected alpha-levels, taking into account that dietary exposures intercorrelate. Finally, abdominal ultrasound is a widely used and reliable imaging technique that yields high sensitivity and specificity for moderate and severe steatosis.53
Nonetheless, there are some limitations that need to be addressed. First, due to the cross-sectional design of this study, it is not possible to draw conclusions on causality. Although reverse causality is unlikely (participants were not aware of having NAFLD when filling in the FFQs), residual confounding may still remain. In particular, participants with diabetes (13%) might have adapted their eating habits due to this comorbidity. Nonetheless, we corrected for this potential confounder in the analyses of the third, metabolic model. Second, part of our study population completed the FFQ 5.5 years prior to abdominal ultrasound (ie, RS III). Because dietary data are known to be stable over time,26 we assumed that dietary habits in this elderly population would be rather constant. This was indeed recently shown in a paper from The Rotterdam Study.54 Nevertheless, study cohort was added as covariate in all regression models and sensitivity analysis was performed excluding the third cohort from all main analyses. The results of this sensitivity were largely similar although no longer statistically significant due to smaller sample size. Third, as with any self-administered questionnaire, data are subject to potential reporter and recall bias. This is reflected in the probable caloric under-reporting in overweight participants.55 However, the 389-item FFQ used in this study has been extensively validated in previous studies.19 20 In addition, unreliable FFQs were excluded. Moreover, we adjusted for total energy intake and thereby accounted for extraneous variation in energy intake and potential measurement error.22 Fourth, in an attempt to avoid bias due to missing data, we performed multiple imputations on our data. Contrary to our expectation, the complete case analyses showed more pronounced associations with NAFLD. Differences between the complete and imputed cases were marginal, but the imputed group had more frequently metabolic disorders. We therefore hypothesise that this somewhat more unhealthy group could have affected the association through a phenomenon in which a relative contribution of a macronutrient to an ‘already higher risk group’ is less pronounced.56 Either way, this could have more likely led to underestimation (rather than overestimation) of the effect that we observed. Finally, in terms of generalisability to younger, non-Caucasian cohorts, results have to be interpreted in the context of a different range of consumption of the various macronutrients by this predominant elderly population.
In conclusion, we found an independent association of high animal protein consumption and NAFLD in an overweight, predominantly aged Caucasian population. The results of this large study add to the current evidence on the importance of dietary composition in NAFLD. In particular, it shifts focus from the carbohydrate and fat debate towards the third, previously underexplored macronutrient, protein. The cause-effect relation and mechanistic pathways of this association remain unanswered for which more studies are needed. Ultimately, we need to understand more about the dietary components that put individuals at risk for NAFLD, before we can make any firm dietary recommendations for the prevention and treatment of NAFLD.
The authors thank the Rotterdam Study participants and staff, in particular, the collaborating general practitioners and pharmacists. The authors are also deeply grateful to Mrs van Wijngaarden (nurse ultrasonographist) for performing the abdominal ultrasonography and transient elastography measurements as well as Jeanne de Vries and Saskia Meijboom from Wageningen University for their assistance in dietary data calculations.
Contributors LJMA: study concept and design, acquisition of data, statistical analysis, analysis and interpretation of data, drafting of the manuscript and finalising the article. JCK-dJ: study concept and design, acquisition of nutritional data, analysis and interpretation of data, statistical analysis, critical revision of the manuscript for important intellectual content and approval of the final article. NSE: imputation procedure and approval of the final article. BJV: study concept and design and approval of the final article. JDS: analysis and interpretation of nutritional data and approval of the final article. RJdK: technical support and approval of final article. MAI: study supervision and approval of the final article. HJM: study concept and design, critical revision of the manuscript for important intellectual content and approval of the final article. JHLA: obtained funding and approval of the final article. OHF: study supervision and approval of the final article. SDM (guarantor): study concept and design, principal investigator of the hepatology department within the Rotterdam Study, analysis and interpretation of data, study supervision, critical revision of the manuscript for important intellectual content and approval of the final article.
Funding The Rotterdam Study is supported by the Erasmus MC University Medical Centre and Erasmus University Rotterdam, the Netherlands Organization for Scientific Research (NWO), the Netherlands Organization for Health Research and Development (ZonMw), the Research Institute for Diseases in the Elderly (RIDE), the Ministry of Education, Culture and Science, the Ministry of Health, Welfare and Sports, the European Commission (DG XII) and by the Municipality of Rotterdam. JCK-dJ received a grant from the Den Dulk-Moermans foundation (Leiden University Fund).
Disclaimer The manuscript’s guarantor affirms that the manuscript is an honest, accurate and transparent account of the study being reported; that no important aspects of the study have been omitted and that any discrepancies from the study as planned have been explained.
Competing interests None declared.
Patient consent Not required.
Ethics approval All cohort participants signed a written informed consent at enrolment. The Rotterdam Study has been approved by the institutional review board (Medical Ethics Committee) of the Erasmus MC University Medical Centre Rotterdam and by the review board of The Netherlands Ministry of Health, Welfare and Sports.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement No additional data available.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.