Meta-analysis: Insulin Sensitizers for the Treatment of Non-alcoholic Steatohepatitis

M. O. Rakoski; A. G. Singal; M. A. M. Rogers; H. Conjeevaram


Aliment Pharmacol Ther. 2010;32(10):1211-1221. 

In This Article


Literature Search

A computer-assisted search was conducted to identify potentially relevant publications in the OVID MEDLINE database on 27 January 2010. The search (1997 to January 2010) was performed using the following exploded (exp) medical subject heading (MeSH), textwords, and truncated text words ($): metformin, thiazolidinedione (exp), pioglitazone, rosiglitazone, insulin sensitiz$, therapeutics (exp), treatment, fatty liver (exp), non-alcoholic fatty liver disease, non-alcoholic steatohepatitis, steatosis, NAFLD, NASH, liver. The preceding terms were combined with appropriate Boolean logic, and this search was then limited to humans and adults. A similar search was also performed in EMBASE, Pubmed and the Cochrane Central Register of Controlled Trials from 1997 to 2010. Additional electronic searches of Digestive Diseases Week (DDW) abstracts, American Association for Study Liver Diseases of (AASLD) meeting abstracts, American College of Gastroenterology (ACG) meeting abstracts and European Association for the Study of the Liver (EASL) meeting abstracts for the period from 2005 to 2009 were performed with combinations of the search terms 'fatty liver' or 'steatohepatitis' or 'NASH' and 'treatment'. A manual search of abstracts from these years was also performed. Manual recursive searches of references from review articles and published randomized controlled trials that met inclusion criteria were also completed. Finally, consultation with expert hepatologists was performed to identify any additional abstracts or unpublished data.

Study Selection Criteria

Potentially relevant studies were selected based on the following inclusion criteria: (i) randomized controlled trials using an insulin sensitizer drug that is currently FDA-approved (metformin, rosiglitazone, pioglitazone); (ii) diagnosis of NASH determined by histology; (iii) liver biopsy performed at beginning and end of clinical trial; and (iv) adult patients. Studies in which NASH was diagnosed based on elevated transaminases or abnormal imaging studies (ultrasound, CT scan, magnetic resonance imaging) without histological confirmation were excluded. Studies that involved weight loss therapies such as orlistat or bariatric surgery were not included because they primarily target weight loss and only secondarily affect insulin resistance.[18]

All randomized controlled trials, regardless of publication status, number of patients randomized, language, or blinding were included. Studies published only as abstracts were included if they had sufficient information on study design, characteristics of participants, interventions and outcomes and/or if an author of the abstract could provide this information. One investigator (MR) reviewed the titles and abstracts of all citations identified by the literature search. Potentially relevant studies were retrieved. Two investigators (MR, AS) independently applied the selection criteria, and any disagreements were resolved by consensus. Agreement between investigators for selection of studies for the meta-analysis was greater than 95%.

Data Extraction

Eligible articles were reviewed in a duplicate, independent manner by two investigators (MR, AS). For each study, the investigators collected the following data: (i) Study: year, location, design, publication status; (ii) Treatment: dose, frequency, duration, active diet and exercise education throughout trial; (iii) Patients: number, mean age, gender, presence of diabetes; (iv) Anthropometrics and laboratory tests (baseline and end of study): BMI, ALT; and (v) Histology (baseline and end of study): steatosis grade, hepatocyte ballooning score, lobular inflammation score and fibrosis stage. If data were not available in the published article, the investigators contacted the corresponding authors for additional information. Any discrepancies in data quantification were resolved by discussion among the investigators.

Primary and Secondary Outcomes

The primary outcomes for this meta-analysis were histological responses to treatment, specifically changes in: (i) steatosis grade (0–3), (ii) hepatocyte ballooning score (0–2), (iii) lobular inflammation score (0–3), or (iv) fibrosis stage (0–4). The secondary outcomes included (i) biochemical response, measured as change in ALT and (ii) anthropometric response, measured as change in body mass index (BMI).

Quality Assessment

Study quality was assessed using the Cochrane Collaboration's tool for assessing risk of bias.[19] As opposed to other assessment tools that utilize scales or check lists, this tool utilizes domain-based evaluation, which focuses on criteria that are more highly associated with internal validity in RCTs. Quality assessment was based on the following domains: (i) randomization, (ii) allocation concealment, (iii) blinding of participants, personnel, and outcome assessors, (iv) completeness of outcome data, (v) unbiased outcome reporting and (vi) lack of other sources of bias. Studies were categorized as high quality if all six domains were well described and low quality if one or more domains were not well described.

Statistical Analysis

The outcomes were calculated as mean difference in scores in the intervention group and were compared with the mean difference in scores in the control group, as described by the following equation: (treatment mean at baseline – treatment mean at the end of the study) – (control mean at baseline – control mean at the end of the study). Therefore, if the histological response to treatment improved to a greater extent in the treatment compared with the control group, the weighted mean difference (WMD) would be positive. Likewise, if the response improved to a greater extent in the control group compared with the treatment group, the WMD would be negative. Difference in means for each outcome for intervention and control group was either directly obtained from the study results or calculated by determining the difference between reported mean values before and after treatment. However, studies did not report the s.d. of the change in means, but reported only the s.d. of the preintervention mean and the s.d. of the postintervention mean. One study provided the s.d. of the change in means for all outcomes upon request.[25] For those studies that reported a P-value for the comparison between pre- and postintervention mean values,[26–28,33] a s.d. for the change in means was derived through back-calculation. For those studies without P-values, t-statistics, or confidence intervals,[30–32] we imputed the s.d. of the change in means, using a modified method by Follmann et al.: s.d.change in means = sq root [(s.d.pre)2 + (− (2(ρ) × s.d.pre ×].[20] This equation estimates the s.d. of the change in means (s.d.change in means) using the s.d. of preintervention mean (s.d.pre) and the s.d. of postintervention mean ( as well the within-participant correlation (ρ) of the outcome measure. We initially assumed a within-participant correlation of zero to derive the largest standard deviation of the change in means, which thereby calculates the most conservative measurement of pooled s.d. for this meta-analysis. A subsequent sensitivity analysis was performed using varied measurements of ρ (ρ = 0.4, ρ = 0.8) to evaluate if varied levels of within-participant correlation altered the statistical significance for each outcome.

Estimates of effect, WMD, were pooled using the DerSimonian and Laird method for a random effects model. For each outcome variable, a forest plot was created that illustrates both study-specific and pooled WMD with 95% confidence intervals. Heterogeneity was assessed using the Cochrane Q statistic and the inconsistency index (I2).[21,22] A chi-square P-value <0.05 or I2 >50% is consistent with the possibility of substantial heterogeneity.[22] Influence analysis, in which one study is removed at a time from the model, was performed to determine if there was a possible undue influence of a single study.[23] Publication bias was graphically examined using a symmetric funnel plot, and then statistically using the test of Egger.[24] A contour-enhanced symmetric funnel plot helps to rule out bias due to small studies that were not published due to unfavourable results. All statistics were computed using stata 11.0 (StataCorp LP, College Station, TX, USA).

Subset analysis was performed for predefined subsets of studies based on (i) type of insulin sensitizer, (ii) exclusion of patients with diabetes, (iii) use of concurrent diet and exercise, (iv) inclusion of patients with benign steatosis in addition to patients with NASH, (v) length of follow-up and (vi) study quality.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: