Table of Contents
The Ethical Review Board of Sweden approved the study and all participants signed a written informed consent prior to inclusion (Dnr: 2019-05111 and 2020-05470).
Metabolic-dysfunction Associated Steatotic Liver Disease (MASLD)
In June of 2023, a multi-society Delphi consensus statement was released proposing a change in nomenclature from NAFLD to Metabolic-dysfunction Associated Steatotic Liver Disease (MASLD)35. This change in nomenclature with its minor modifications to the disease criteria occurred late in the analysis phase of this study. To stay true to the primary name of the trial as well as to further prespecified subgroup analyses specified below, we chose to keep the initial name of NAFLD for this manuscript, with the caveat that nearly 100% of all individuals with NAFLD also meet the criteria for MASLD in Sweden36. In the current study, the overlap was 100%.
Study design
The NAFLDiet trial was a single-site, three-arm parallel-group designed study that randomized participants in a 1:1:1 allocation ratio to follow either a LCPUFA diet, a HND or UC, in line with the Nordic Nutrition Recommendations (NNR) 2012 (ClinicalTrials.gov Identifier: NCT04527965). The NAFLDiet trial was conducted in Uppsala, Sweden between August 2020 and December 2022. The first participant was enrolled 11th of August 2020 and the last participant 15th of December 2021.
Participants were predominantly recruited from a local diabetes register (ANDiU) and a large population-based cohort (EpiHealth), but also through web-based advertisement. Men and women (self-reported) were eligible to participate if they met the following criteria: 30–75 years of age, Body Mass Index (BMI) 25–40 kg/m2, T2D with a duration ≤10 years with no insulin treatment or prediabetes (defined as a fasting plasma glucose ≥5.6 mmol/L or an HbA1c ≥ 39 mmol/mol in accordance with guidelines from American Diabetes Association 2019) and no diagnosed cardiovascular disease (CVD) in the last two years prior to screening. Exclusion criteria included self-reported alcohol intake >20 g/day, contraindications for magnetic resonance imaging (MRI) assessment, unwillingness to follow a new prescribed diet for a year, ≥10% diet-induced weight loss the preceding three months of screening, malignant diseases, severe kidney or liver disease, heart failure or other severe CVD. A total of n = 222 individuals were scheduled for a physical screening visit, of which n = 150 were included and randomized (Fig. 1).
Randomization (stratified by sex and T2D status) was performed by a researcher not involved in the execution of the study (i.e., no contact with participants and no involvement in primary or secondary outcome ascertainments), using a computerized random-number generator. Knowledge of the allocation sequence was restricted to this researcher. To avoid spill-over effects from the diets, four pairs of participants who lived together were co-randomized. Personnel responsible for assessing study outcomes (including dietary adherence) as well as care providers were blinded to the assigned diets of the participants. Neither the participants nor the study coordinator (M.F.) were blinded to the assigned diets. Participants did not know of their assigned diet until they had completed all measurements from the fasting study baseline visit.
Diets
At baseline, participants were provided with detailed written and orally presented dietary guidelines tailored to their assigned diet.
LCPUFA
The LCPUFA group was instructed to limit their intake of carbohydrates to <30 E%, but maintain whole-grain and fiber-rich carbohydrates, and increase their intake of fat to >50 E%, (focusing on increasing omega-6 PUFA and limiting SFA), while keeping protein at 20 E% (focusing on plant-based food sources). To achieve this, ten food-based dietary recommendations were advised, of which four were emphasized; consume at least two table spoons (tbsp.) of sunflower oil per day, two tbsp. of seeds (primarily pumpkin- and sunflower seeds) per day and a handful of nuts (40 × g) (primarily walnuts, pecan nuts and Brazil nuts) per day. Participants were also guided (through both an information leaflet and a recipe book) on how to replace carbohydrate sources such as pasta, rice, bread and potatoes with other food sources such as bean pasta, cauliflower rice, low-carbohydrate bread rich in seeds and nuts, and root vegetables such as carrots and parsnips.
HND
The HND group was instructed to limit their intake of fat to 25–30 E% (focusing on PUFA and MUFA) and increase their intake of carbohydrates to 50–55 E% (focusing on increasing whole-grain, fiber-rich carbohydrates and limiting refined carbohydrate sources), while keeping protein at 20 E%. To achieve this, ten food-based dietary recommendations were provided, of which four were emphasized; participants were told to consume at least two portions of whole-grains (primarily oats and rye) per day, two slices of whole-grain bread (with flour made from rye and/or oats) and to focus on foods grown or produced in Sweden and other Nordic countries, such as apples, pears, blueberries, raspberries, cabbage, almonds, salmon, mackerel and herring. Participants were also guided (through both an information leaflet and a recipe book) on how to limit fat intake by replacing full-fat products with low-fat products (e.g., milk, fermented milk and cheese) and considering the amount of salad dressing, cooking oil and bread spread used. The main fat source that was emphasized in the HND group was rapeseed oil.
UC
The UC group received information on how to follow the current dietary guidelines laid out in the NNR 2012 edition. Protein intake was kept at 20 E%. To partly mask participants from knowing that they had been randomized to the usual care group, extra emphasis was put on increasing fruits and vegetables to 600 grams per day instead of the current recommendation of 500 grams per day. No clear emphasis was placed on using local or Nordic foods in this group. A wider variety of healthy foods were advised, e.g., in contrast to the HND, olive oil could be used instead of rapeseed oil or sunflower oil, and all types of fruit were advised rather than only Nordic types such as apples and pears. The UC group also received an information leaflet and a recipe book.
To enhance diet adherence, study participants received key food items on a monthly to bimonthly basis (once every month the first 7 months followed by once every two months the remaining 5 months). For the LCPUFA group, these foods included sunflower oil, walnuts, cashew nuts, pumpkin seeds and sunflower seeds. For the HND group, these foods included oats, oat bran, oat rice, whole-grain muesli, crisp bread, low-fat margarine (based on rapeseed and sunflower oil), raspberries, beans, lentils and almonds. For the UC group, key food items included crisp bread, whole-grain cereals based on oats, carrots, frozen mango and frozen peas.
All diets emphasized increasing the intake of fruits and vegetables, replacing whole-fat dairy with low-fat dairy and limiting red and processed red meat as well as sugar-sweetened beverages, pastries and other energy-dense sugar/fat-rich snacks. Diets were provided ad libitum; thus, weight loss was not explicitly targeted. More details of the diets can be found in Supplementary Table 1. To mitigate introducing unwanted co-interventions that may influence the size of the effect of both primary and secondary outcomes, participants were instructed to make no major changes in physical activity, alcohol consumption habits or dietary supplements use. However, due to ethical reasons, participants were allowed to make changes to their medications.
Self-reported and biomarker based dietary adherence and alcohol intake
Dietary adherence was partly assessed using 4 days (3 weekdays and 1 weekend day) weighed food diaries (WFD) at baseline, 6 months and 12 months. Participants were instructed to weigh all their foods and drinks (excluding water) and report the weights in grams. If weighing the food and/or drinks was not feasible, participants were instructed to report the volume of foods and drinks consumed using household measures. The WFD were processed and analyzed in the commercially available software Dietist Net linked to the Swedish National Food Agency (SNFA) food database by a nutritionist not involved in the design or execution of the trial.
Fatty acids in plasma were analyzed and used as objective biomarkers of fat intake. Specifically, the major dietary PUFA, linoleic acid (18:2n-6) in plasma phospholipids (PL) was used as a valid biomarker of PUFA intake from vegetable fats, and omega-3 eicosapentaenoic acid (EPA) (20:5n-3) and omega-3 docosahexaenoic acid (DHA) (22:6n-3) in PL were used as biomarkers of omega-3 PUFA from seafood mainly. Also, palmitoleic acid (16:1n-7) in plasma triacylglycerols (TAG) was used as a marker of carbohydrate-induced DNL37,38,39. Plasma lipids were extracted from fasting blood samples using the Folch method and lipid fractions were subsequently separated using solid phase extraction (SPE)40. Fatty acids were methylated and analyzed using gas chromatography (GC), and given as proportions of the total fatty acids analyzed (peak area%).
Plasma alkylresorcinols (AR), biomarkers of whole-grain wheat and rye intake were analyzed by an LC-MS/MS-based method as reported previously41. The within and between batch CVs were 15%. Total AR and the ratio between the homologues 17:0 and 21:0 (17:0/21:0) were measured to assess adherence to whole-grain intake from rye and wheat and the proportion of rye to wheat in the diet, respectively42.
Given the influence of alcohol on liver fat, triglycerides and glycemic control, alcohol intake was assessed by WFD as well as by phosphatidylethanol (PEth), a specific and sensitive biomarker of alcohol intake as measured in plasma at the Uppsala University Hospital clinical chemistry laboratory.
Primary outcome
For quantification of liver fat, a single breath-hold 6-echo 3D water-fat MRI protocol (IDEAL-IQ) was acquired using a surface coil (UAA). The vendor implemented water-fat and fat fraction reconstructions were used. Scan parameters were: TR = 6.27 ms, TE = 2.45 ms, flip angle 3 deg, nsa 0.73, slice thickness 5 mm, FOV 384 × 384 mm, 90% Phase FOV. Axial orientation was used and in-slice voxel dimensions were 1.5 × 1.5 mm and 5 mm slice thickness was used. Number of slices exported from each scan was 28. Liver mean proton density fat fraction (PDFF) was quantified using an automated segmentation approach. Reference liver segmentations were first created manually by delineations in the water signal images from 26 randomly selected subjects. A UNet++ 2D convolutional neural network was then trained to perform the segmentations using the water-signal images using the Dice loss function43. A test set of 8% of the data was set aside. Ten-fold cross validation resulted in a segmentation accuracy, in terms of mean Dices scores of 0.981. A mean Dice score of 0.983 was obtained on the test set. The trained model was then applied to all images. Post processing was applied using 3D erosion and filtering of largest connected object in 3D. This to reduce the effect of eventual segmentation errors. All segmentation masks were visually quality controlled. Outlier voxel fat fraction values were also excluded outside ±3 SDs from the liver mean fat fraction. The mean liver proton density fat fraction was uses as the measurement of liver fat content44. Outcome assessors were blinded to participants’ diet allocations. For subgroup analyses, NAFLD was defined as a liver fat content exceeding 5.6%.
Secondary and exploratory outcomes
Fasting plasma concentrations of glucose, HbA1c, total cholesterol, LDL-cholesterol, HDL-cholesterol, apoA1, apoB, triglycerides, ASAT, ALAT, GGT, platelet counts and C-reactive protein (CRP) and fasting serum concentrations of insulin were measured by routine laboratory methods at Uppsala University Hospital. FIB-4 was calculated as \(({age\; x\; ASAT})/({platelet\; counts\; x\; ALAT})\) and HOMA-IR was calculated as \(({insulin\; x\; glucose})/22.5\). Systolic (SBP) and diastolic (DBP) blood pressure were assessed in a sitting position at the research clinic by a specialized research nurse after five minutes of rest using an automated blood pressure device (Omron). The values of three consecutive measurements (allowing for 1 min of rest between) were averaged and reported. Body weight and height were measured at the research clinic after an overnight fast. Height was measured using a stadiometer to the nearest 0.5 cm. Weight was measured in light clothing using bioelectrical impedance analysis (BIA) (Tanita). BMI was calculated as \({weight}({kg})/{height}(m)2\).
Statistical analysis
Sample size calculation was based on Lehr´s formula for the comparison between groups, assuming equal treatment effects for the two experimental diets. A sample size of n = 37 in each group was estimated to detect a 2 (SD: ± 3) percentage unit difference in liver fat between the experimental groups and the UC group, with significance level (α) of 0.05 and power (1-β) of 0.80. To achieve the desired power for both the intention-to-treat and the per-protocol analyses, and allowing a 25% dropout rate, we included 50 participants in each group. As NAFLD is diagnosed by a liver fat content exceeding 5.6%, a 2% absolute difference was determined clinically meaningful. A 3% SD was determined based on previous trials in healthy individuals or individuals with overweight or obesity5,6.
An intention-to-treat (ITT) effect was determined a priori to be the primary effect of interest. All individuals who were randomized and were informed of their assigned diet were included in the ITT-population (Fig. 1). A general linear model (GLM) with diet group as a fixed factor, the outcome as dependent variable (difference between month 12 and baseline) and the following baseline variables as covariates: value of the outcome at baseline, sex (man/woman) and T2D diagnosis (yes/no) were included. Multivariate imputation using chained equations (MICE) (n = 20 imputations with predictive mean matching (PMM) as imputation method) was used to account for missing outcome and baseline data, assuming the data was missing at random (MAR). MICE was performed using the mice package in R45. The amount of missing data ranged from 5% to 7% for all outcomes except for liver fat that had 14% missing data (Supplementary Table 2). The imputation model included diet group, BMI, age, sex, T2D diagnosis, baseline value of the outcome, and the outcome as variables. Pooling of parameter estimates from each imputed dataset was performed using Rubin´s rules. When the pooled F-parameter estimate of the GLM was statistically significant, linear models for pairwise comparisons were performed. This strategy preserves the type 1 error rate at the nominal level for comparisons of three groups46. Between-group differences were expressed as estimated marginal means (EMM) with corresponding 95% confidence intervals (CI). Assumptions of the GLM were checked using residual plots. The Shapiro–Wilk test for residuals was used to examine whether data was normally distributed, with a test statistic W > 0.95 indicative of a Gaussian distribution.
When the assumptions of the GLM were not satisfied, data were analyzed using Willett´s residual method47, which is a non-parametric test for group differences with adjustment for covariates, and was performed as follows. A similar GLM as described above was fitted to the data but with diet group omitted from the model. Residuals from the GLM were saved and subsequently used as the dependent variable in a Kruskal–Wallis test, with diet group included as factor. Estimated marginal median differences between groups with corresponding 95% CI and p-values were retrieved via bootstrapping (n = 10,000 bootstraps, n = 20 imputations using MICE) with the use of the bootImpute package in R48 if the Kruskal–Wallis test was statistically significant.
A per-protocol analysis, including all participants who completed the intervention and who had data on the outcome of interest was determined to be secondary to the ITT-analysis. Similar statistical methods were employed for the estimation of the per-protocol effect except for when the assumptions of the GLM were not satisfied, where the Hodges–Lehman estimator was used to retrieve estimated marginal median differences with corresponding 95% CI.
Prespecified subgroup analyses (by sex, T2D diagnosis, PNPLA3 I148M genotype, and NAFLD status) were performed for the estimation of both the ITT-effect and the PP-effect for the following outcomes: liver fat, HbA1c, LDL-C, total cholesterol, HDL-C, triglycerides, apoB and apoA1. Similar models were specified as for the full population with the exception that the stratifying variable was excluded from the model. Multiple imputation and bootstrapping were performed separately for each subgroup. Due to missing information on baseline liver fat content in n = 6 individuals, NAFLD status was first imputed, followed by stratification. The robustness of our findings was explored in sensitivity analyses by excluding participants who were co-randomized as pairs (n = 4 pairs) and by including values of BMI at month 6 and month 12 in the MICE model. Sensitivity analyses were not prespecified in the statistical analysis plan (SAP) (Supplementary Note 1).
Lastly, as an additional post-hoc analysis, a causal mediation analysis (CMA) for our primary outcome was conducted using the regmedint package in R with weight change from 0 to 12 months included as the mediator49,50,51. Proportion mediated (PM) by weight change was calculated by dividing the natural indirect effect (NIE) by the total causal effect (TE). A detailed description of the methodology underlying the CMA can be found in Supplementary Material.
The original SAP can be found at ClinicalTrials.gov (ClinicalTrials.gov Identifier: NCT04527965). We have modified the statistical analysis compared to the SAP to allow comparisons between all three treatment groups as described above. Other minor modifications to the SAP are described in supplementary material. A p-value < 0.05 (two-sided) was determined to be statistically significant. R (R Core Team, Vienna, Austria) version 4.2.3 and IBM SPSS Statistics version 28.0.1.0 (142) were used to analyze the data.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
