Dietary factors that influence NAFLD have become a focus of attention. In particular, recent concerns have been raised regarding the role of dietary fructose in inducing NAFLD.4, 5, 6, 7 Animal models featuring extreme levels of fructose exposure8, 9, 10, 11 and poor-quality observational studies12, 13, 14 have been used to underpin this hypothesis. In the absence of consistent clinical evidence, it is unclear whether fructose at typical levels of exposure induces NAFLD. To determine the effect of fructose on markers of NAFLD in humans, we conducted a systematic review and meta-analysis of available controlled feeding trials.
We followed the Cochrane Handbook for Systematic Reviews of Interventions and the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines. The review protocol is available at ClinicalTrials.gov (registration number: NCT01363791).
We searched MEDLINE, EMBASE, CINAHL and the Cochrane Library databases through 3 September 2013 for relevant articles. Supplementary Appendix Table 1 shows the full search term used in this study. Manual searches supplemented the electronic search strategy. No restrictions were placed on language. We included controlled trials investigating the effect of oral fructose on markers of NAFLD. A comparison was considered isocaloric when the carbohydrate comparator was exchanged for an equal amount of fructose. If the trial involved overfeeding of fructose so that the fructose provided excess energy resulting in a positive energy balance, then the comparison was still considered isocaloric as long as the carbohydrate comparator was matched for the excess energy resulting in the same positive energy balance. A comparison was considered hypercaloric when a control diet was supplemented with excess energy from fructose compared with the same control diet alone without the excess energy. Trials that involved a follow-up of <7 days follow-up, administered intravenous fructose, lacked a control diet or did not provide suitable endpoint data were excluded.
Two reviewers (SC, AIC) independently reviewed and extracted relevant data from each report. The quality of each study was assessed using the Heyland methodological quality score (MQS). Disagreements were reconciled by consensus. Mean±s.d. differences between fructose and control arms were extracted as the main end points. In those trials where the data were included in figures and not provided numerically, we used software program Plot Digitizer (http://plotdigitizer.sourceforge.net/) to extract the data. Additional information was requested from the authors of all included trials.
All authors had access to the study data and reviewed and approved the final manuscript.
Data analyses were conducted using Review Manager (RevMan) version 5.1.6 (Copenhagen, Denmark) for primary analyses and Stata (version 12, College Station, TX, USA) for subgroup analyses. Separate analyses were conducted for the isocaloric and hypercaloric trials using the generic inverse variance method with random effects weighting. Data were expressed as standardized mean differences with 95% confidence intervals (CIs) for intrahepatocellular lipid (IHCL) and mean differences (MD) with 95% CIs for alanine aminotransferase (ALT).
Trials that did not report SE values had these computed from the available statistics using standard formulae. To generate SE for included crossover trials, we assumed a paired analyses as described by Elbourne. If insufficient data were available for computations in crossover trials, SE values were imputed using the pooled correlation coefficient between baseline and end-of-study values derived from a meta-analysis of trials reporting sufficient data or assuming a conservative correlation coefficient of 0.5 with sensitivity analyses at 0.25 and 0.75.
Inter-trial heterogeneity was assessed by the Cochran Q statistic with α<0.10 considered significant, and quantified by the I statistic, where I⩾50% indicates substantial heterogeneity. Sources of heterogeneity were investigated by sensitivity analyses in which each individual study was removed from the analysis and through a priori subgroup analyses by comparator, baseline values, fructose form, follow-up, MQS, randomization, design and energy balance. Meta-regression analyses assessed the significance of subgroup effects. Publication bias was evaluated via visual inspection of funnel plots and Egger and Begg tests.