Literature Search
We searched MEDLINE, the Cochrane Library, and EMBASE without language restrictions for original research articles, systematic reviews, and meta-analyses to identify all available literature on omega-3 fatty acids and cardiovascular disease published from January 1966 through September 2008. We combined search terms for omega-3 fatty acids ("omega-3 fatty acids" OR "fish oil" OR "marine oil" OR "dietary therapy") with those for mortality or cardiovascular disease ("mortality" OR "cardiovascular disease" OR "heart disease" OR "CAD" OR "MI" OR "UA" OR "coronary angiography" OR "coronary restenosis"). References of relevant articles were hand-searched for additional studies.
Inclusion criteria required each study to 1) be a comparative randomized trial involving human participants with an active treatment arm using omega-3 fatty acids with usual diet; 2) involve a high-risk population with known cardiovascular disease or diabetes; and 3) report at least 1 of the following 2 outcomes: total mortality or coronary artery restenosis following angioplasty. We excluded non-randomized studies, those involving children or animals, and studies in which the omega-3/fish supplement dosage was unspecified. We also excluded studies in which patients were randomized to dietary advice that included a non-quantifiable intervention of simply increasing fish consumption [
21]. The Quality of Reporting of Meta-analyses of Randomized Controlled Trials (QUORUM) guidelines were followed throughout this meta-analysis [
22] [See Additional file
1 - QUORUM checklist].
One author performed the literature search, and data extraction was independently conducted by at least two individuals. The following information was extracted: publication details, timing of study, duration of follow-up, randomization method, blinding (of participants, investigators, and outcome assessors), omega-3 dosage, dropouts, mean age of participants, primary outcome (total mortality), secondary cardiovascular outcomes (restenosis, sudden death, cardiac death, non-fatal myocardial infarction (MI), congestive heart failure (CHF), arrhythmias, implantable cardioverter defibrillator (ICD) shocks, and stroke), the proportion of patients who discontinued treatment, side effects (gastro-intestinal (GI) side effects, bleeding, and malignancies), and adherence. Any disagreements in the collected data was resolved by consensus or, when necessary, upon consultation with a third reviewer. The authors of the original publications were contacted to obtain missing data and resolve ambiguities (n = 16), although we made no attempt to contact authors of articles published before 1995.
Quality assessment of the individual trials was performed using the Jadad scale [
23] [See Additional file
2-Quality assessment of included trials].
Statistical Analysis
To estimate the risk of all-cause mortality, data were analyzed according to the intention-to-treat principal; the denominator was the number of participants randomized to each group, and the numerator was the number of deaths reported during the follow-up period. In our restenosis analysis, the denominator was the number of participants undergoing follow-up coronary angiography, and the numerator was those with restenosis. In most studies, restenosis was defined as the loss of luminal diameter of at least 50% [
24‐
29]. One study defined it as a loss of 70% [
30], and 2 studies defined coronary restenosis as at least 50% stenosis at follow-up [
31,
32]. Three trials used multiple definitions, including loss of luminal diameter of at least 50% and at least 50% stenosis at follow-up [
33‐
35]. The remaining two studies used unique restenosis definitions; one used a panel of blinded cardiologists to assess changes in progression and regression of CAD [
36], and the other defined restenosis as lumens with greater than 20% obstruction [
37]. Where possible, cardiovascular event and safety data were analyzed using an intention-to-treat approach. However, many restenosis studies presented data only for those who underwent follow-up coronary angiography.
For each study, we estimated the risk ratio (RR) comparing intervention and control groups. For studies with zero outcomes in either group, we added 0.5 to all cells of the 2-by-2 table. For the primary outcomes of mortality and main secondary analysis of restenosis, we fit meta-regression models to investigate if study-level covariates explained any of the heterogeneity in RRs across studies. We estimated the median and corresponding 95% credible interval (CrI), the Bayesian analogue for confidence intervals, for each coefficient in these models. In addition, we also estimated the probability that the coefficient was greater than 0. The covariates used in the meta-regression were median follow-up time (months), sample size, high dosage of the intervention, high quality (on the Jadad scale), high adherence rate, and high percentage of previous MI. Cut-offs for defining dichotomous covariates were determined by the median value across studies. For the all-cause mortality meta-analysis, subgroups were defined by sample size (>322 patients), study quality (>3), dosage of omega-3 (>3.3 g/day), adherence (>84%), and history of previous MI (>50%). For the restenosis meta-analysis, subgroups were defined by study quality (>3), sample size (>233 patients), dosage of omega-3 (>5.04 g/day), adherence (>85%), and previous MI (>48%).
We carried out meta-analyses to pool RRs across all studies [
38]. Separate models were created for each of the primary and secondary outcomes and for safety data. When there was greater than 90% probability that a covariate was associated with the relative risk (i.e., >90% probability that the meta-regression coefficient was different from 0), we carried out separate meta-analyses within subgroups defined by the covariate. Funnel plots were employed to assess publication bias.
All analyses were carried out using WinBUGS and R software programs. Low information prior distributions (mean 0 and a standard deviation of 100) were used for all parameters. The between-study standard deviation in the log (RR) was assumed to follow a uniform distribution over the range from 0 to 5.