Materials and Methods
The protocol for this review was registered in advance with PROSPERO (International Prospective Register of Systematic Reviews, CRD42019135559). We conducted and reported this systematic review according to the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) checklist and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.
Inclusion Criteria, Exclusion Criteria and Study Outcome
Studies were included if they were a randomized controlled trial (RCT), cross-sectional study or cohort study, including both prospective or retrospective designs, which reported fatty liver events in adult patients (more than 18 years old) taking tamoxifen. The study participants were restricted to breast cancer patients without restrictions for comorbid conditions. Fatty liver was defined based on radiological imaging [ultrasound sonography (USG), computed tomography (CT)], histology or noninvasive serum markers using serum liver enzymes. Studies that met the following three inclusion criteria were selected for the analysis: (a) the study population consisted of breast cancer patients taking tamoxifen, (b) the association between tamoxifen and the development of fatty liver was assessed and (c) the outcome of interest was the prevalence or the incidence of fatty liver. The exclusion criteria were: (a) case reports, (b) case series for patients with total numbers less than five, (c) review articles, (d) cell studies or animal studies, (e) studies not showing the incidence of fatty liver and (f) studies including patients taking tamoxifen for prophylactic use without breast cancer.
The primary outcome of this study was the prevalence and incidence of fatty liver in patients taking tamoxifen. The secondary outcomes were comparisons of fatty liver incidence according to basal characteristics and the derivation of risk factors for fatty liver induced by tamoxifen.
We searched synonymous terms and used these terms to develop the search strategy. The keywords used in the Patient/Problem, Intervention, Comparison and Outcome (PICO) model are shown in the Supplementary Material. We searched PubMed (Medline), EMBASE and OVID Medline, the Cochrane Library, KoreaMed, MedRIC and KSI KISS to identify studies published in English between 1 January 1987 and 31 May 2018. The search strategies and search results of each database are shown in the Supplementary Material. We excluded studies published before 1987 since the first case report of fatty liver due to tamoxifen was reported in 1987. All these search processes were conducted by a professional librarian (EAJ).
The search words were breast cancer-related index words ("breast neoplasms" OR "breast tumor" OR "breast carcinoma" OR "breast cancer" OR "mammary neoplasm" OR "mammary cancer" OR "mammary carcinoma" OR "mammary tumor" OR "breast malignant neoplasm"), tamoxifen-related index words ("tamoxifen" OR "antineoplastic agents, hormonal" OR "nolvadex" OR "tomaxithen" OR "zitazonium" OR "soltamox" OR "crisafeno") and non-alcoholic fatty liver disease-related index words ("non-alcoholic fatty liver" OR "NAFLD" OR "non-alcoholic steatohepatitis" OR "visceral steatosis" OR "liver steatosis"). Other details are shown in the Supplementary Material.
Study Selection and Data Extraction
The titles and abstracts were independently screened by two authors. Two reviewers (JJY, CBL) independently screened the relevant full-text articles. If any discrepancy between the two reviewers was found, it was resolved by SGK after discussion. In addition, two researchers independently screened all the other remnant stages, including full-text screening and the risk of bias assessment. Two reviewers (JJY, BL) extracted the study characteristics and the results and recorded these data in a standard form with any discrepancy resolved by YSK and SGK.
Methodological Quality and Risk of Bias Assessment
We used two tools to determine the risk of bias, the Cochrane risk of bias tool and the Suggested Quality Scoring System, which was proposed in 2008 by Doi to conduct quality-effects analysis. This information is provided in the Supplementary Material. If any discrepancy occurred, it was resolved by other authors (YSK and SGK) after a discussion. The overall result of the risk of bias is shown in the Supplementary Material.
We derived the pooled prevalence using a random-effects model and a quality-effects (QE) model using the eight cross-sectional studies. We used the metaphor function from R package. The incidence, person-time and the number of fatty liver events among the patients treated with or without tamoxifen were calculated from the cohort studies or RCTs. The incidence rate ratios were polled by the DerSimonian–Laird estimator for τ 2 with an inverse variance weight after log-transforming the study-specific incidence rate per 100 person-years. The risk factors were recorded as HRs with 95% CIs on a log-scale. They were weighted by the inverse of their corresponding variances to obtain pooled estimates.
We fitted the random-effects meta-analysis to account for the moderate to substantial heterogeneity in the prevalence and incidence among the studies according to the Cochrane collaboration. Additionally, a quality-effects model was applied to avoid confounding effects in the meta-analyses of the observational studies, using the quality score computed based on the criteria by Doi et al. We scored the quality of each study based on Doi's criteria and analysed the prevalence, incidence and incidence rate ratio using the metaXL program. To assess the effect of individual studies on the pooled estimates, a sensitivity analysis was conducted by re-estimation after omitting each study, one at a time. Publication bias was assessed using a funnel plot.
We used fixed-effects or random-effects meta-analysis, depending on the heterogeneity between studies. To explain the heterogeneity between the studies, subgroup analyses were conducted on the geographic region (eastern and western) and the fatty liver diagnostic method. Meta-regression was conducted to examine the influence of other factors on the clinical outcomes.
Statistical analyses were performed using RevMan 5 (Cochrane library) to assess the risk of bias and MetaXL (version 5.3; EpiGear International) for the quality-effects model. The overall meta-analyses were performed using the meta package in R (version 3.6.1; The R Foundation for Statistical Computing).
Liver International. 2020;40(6):1344-1355. © 2020 Blackwell Publishing