This study was undertaken according to the methodology of MOOSE (Meta-analysis of Observational Studies in Epidemiology) statement (Stroup et al., 2000). We thoroughly searched all published studies in the Embase, PubMed, China National Knowledge Infrastructure (CNKI) and Wanfang databases up to August 2, 2017. The included articles were limited to Chinese and English language. The following keywords were used for searching: “apolipoprotein E” OR “APOE” AND “polymorphism” OR “single nucleotide polymorphism” OR “SNP” OR “variant” OR “variation” AND “coronary artery disease” OR “coronary heart disease” OR “CAD” OR “CHD” OR “atherosclerosis” OR “myocardial infarction” OR “myocardial infarct” OR “heart attack” OR “MI” AND “type 2 diabetes” OR “non-insulin dependent diabetes mellitus” OR “diabetes mellitus, type 2” OR “diabetes, type 2” OR “diabetes mellitus, non-insulin dependent” The Chinese databases were searched using the equivalent Chinese terms. In addition, hand searches for all related articles were performed. The detailed search strategies are presented in Supplementary Table 1.

The first two investigators independently accessed the eligibility of the studies by screening the title, abstract and full-text, based on the inclusion and exclusion criteria. The inclusion criteria for all studies were as follows: (1) study on the associations between APOE ε2/ε3/ε4 polymorphisms and CAD in patients with T2DM, regardless of sample size. (2) case-control design. (3) detailed data for the APOE alleles or genotype distribution in case and control groups to estimate odds ratio (OR) with 95% confidence interval (CI). Exclusion criteria: (1) duplication of previous data; (2) review, comment and editorial; (3) no sufficient genotype data. Any dispute about the eligibility of an article was resolved by discussion.

The data was drawn out based on a standard protocol. The following information was carefully extracted from all eligible publications independently by two authors (JQL and HR) using a standardized form: last name of first author, year of publication, study country, sample size in cases and controls, methods of genotyping, number genotypes and alleles. If similar data sets presented in different articles by the same research group, the data would be adopted only once. The collected data were compared, and possible disagreements were discussed by the authors and resolved with consensus.

The study quality was independently assessed by two reviewers. Quality assessment of genetic associations between APOE ε2/ε3/ε4 polymorphisms and CAD in patients with T2DM is described in the Supplementary Table 2. The scores were adjusted according to the criteria developed for meta-analysis of molecular association studies by Thakkinstian et al. (2005). The total scores ranged from 0 to 13, with 13 representing the highest quality.

All the statistical analysis in this study was performed using Stata 12.0 (StataCorp, College Station, TX). Hardy-Weinberg equilibrium was performed in control groups using the chi-square test. The combined OR and 95%CI were calculated to evaluate the strength of the association between the APOE ε2/ε3/ε4 polymorphisms and risk of CAD in T2DM subjects. The pooled ORs were, respectively, performed for nine genetic models (ε2/ε2 vs. ε3/ε3; ε2/ε3 vs. ε3/ε3; ε2/ε4 vs. ε3/ε3; ε3/ε4 vs. ε3/ε3; ε4/ε4 vs. ε3/ε3; ε2 allele vs. ε3 allele; ε4 allele vs. ε3 allele; ε2/ε2+ε2/ε3 vs. ε3/ε3; ε4/ε4+ε3/ε4 vs. ε3/ε3). The statistically significant level of the combined OR was determined by the Z-test with P < 0.05. Heterogeneity between studies was calculated by using the Cochran's Q-test and Higgins I² index. In the absence of between-study heterogeneity (I² < 50%), the fixed effect model (Mantel–Haenszel method) was chosen to calculate the pooled estimates. Otherwise, random effect model (DerSimonian and Laird method) would be adopted if the I² > 50% (Higgins et al., 2003). Subgroup analysis was performed according to the source of patients (Chinese and non-Chinese). Galbraith plot analysis and sensitivity analysis were conducted to detect whether there were outliers that could be the potential sources of heterogeneity between studies when heterogeneity was moderately large. Publication bias was evaluated by Begg's funnel plot and Egger's regression test (Begg and Mazumdar, 1994; Egger et al., 1997). If there is evidence of significant publication bias, the trim and fill method was performed to assess the potential influence of publication bias (Duval and Tweedie, 2000).

As depicted in Figure 1, a total of 222 articles were obtained by online search, and 2 articles were included by manual search. After removing duplicates, 175 articles were included. After screening title and abstract, 115 articles were excluded. As a result, 13 articles (Zheng et al., 1998; Zhang et al., 2000, 2008; Pan et al., 2002; Guo et al., 2007; Chaaba et al., 2008; Izar et al., 2009; Shi et al., 2009; Vaisi-Raygani et al., 2010; Al-Majed et al., 2011; Chaudhary et al., 2012; Halim et al., 2012; Hong et al., 2017) were eligible for the meta-analysis. The characteristics of the included articles are summarized in Table 1. The included studies were conducted in several countries including China, Brazil, Thailand, Egypt, Iran, Kuwait, and Tunisia. All studies were performed in a case-control design and the sample sizes varied from 70 to 990. The quality score of the included studies ranged from 5 to 12 (mean: 9.69) out of a maximal score of 13.

The main results of this meta-analysis for the association between APOE ε2/ε3/ε4 polymorphisms and the risk of CAD in T2DM patients are presented in Table 2. There was significant association in three genetic models which demonstrate, ε4 mutation contributed to an increased risk of CAD in patients with T2DM (Figure 2). The pooled results for the three genetic models in the overall analysis were as follows: for ε3/ε4 vs. ε3/ε3: OR = 1.69, 95% CI = 1.38–2.08, P < 0.001; for ε4/ε4 vs. ε3/ε3: OR = 2.72, 95% CI = 1.61–4.60, P < 0.001; for ε4/ε4+ε3/ε4 vs. ε3/ε3: OR = 1.83, 95% CI = 1.52–2.22, P < 0.001. In contrast, the ε2 variation had null association with this disease (Figure 3). No significant association in the overall analysis was found in genetic model of ε2/ε2 vs. ε3/ε3 (OR = 1.67, 95% CI = 0.90–3.09, P = 0.104); ε2/ε3 vs. ε3/ε3 (OR = 1.18, 95% CI = 0.93–1.51, P = 0.175); ε2/ε4 vs. ε3/ε3 (OR = 1.20, 95% CI = 0.78–1.84, P = 0.405); ε2/ε2+ε2/ε3 vs. ε3/ε3 (OR = 1.26, 95% CI = 0.88–1.82, P = 0.212). In addition, the genetic models of allele-based contrasts in the overall analysis also revealed a statistically significant pooled estimates for ε4 allele vs. ε3 allele (OR = 1.64, 95% CI = 1.40–1.94, P < 0.001) but not for ε2 allele vs. ε3 allele (OR = 1.34, 95% CI = 0.98–1.84, P = 0.07).

In the subgroup analysis according to the source of patients (Chinese and non-Chinese), the pooled ORs of all genetic models except the ε2/ε4 vs. ε3/ε3 model were consistent with the results in the overall population. In the Chinese population, the ε2/ε4 genotype increased the risk of CAD in patients with T2DM (OR = 2.17, 95% CI = 1.10–4.28, P = 0.026).

As shown in Table 2, there was moderate between-study heterogeneity in the genetic model of ε2 allele vs. ε3 allele (P_{heterogeneity} = 0.054, I² = 44.70%) and ε2/ε2+ε2/ε3 vs. ε3/ε3 (P_{heterogeneity} = 0.071, I² = 39.50%) in the overall analysis. However, no significant heterogeneity was found in other genetic models (for ε2/ε2 vs. ε3/ε3: P_{heterogeneity} = 0.532, I² = 0%; for ε2/ε3 vs. ε3/ε3: P_{heterogeneity} = 0.151, I² = 30.10%; for ε2/ε4 vs. ε3/ε3: P_{heterogeneity} = 0.493, I² = 0%; for ε3/ε4 vs. ε3/ε3: P_{heterogeneity} = 0.312, I² = 13.90%; for ε4/ε4 vs. ε3/ε3: P_{heterogeneity} = 0.807, I² = 0%; for ε4 allele vs. ε3 allele: P_{heterogeneity} = 0.284, I² = 16.80%; for ε4/ε4+ε3/ε4 vs. ε3/ε3: P_{heterogeneity} = 0.384, I² = 6.20%). The heterogeneity analysis results indicated that the pooled results of this meta-analysis in most genetic models were statistically steady and robust. In addition, subgroup analysis indicated that there was no heterogeneity under all nine genetic models in the Chinese population.

There was evidence of moderately large between-study heterogeneity in the genetic model of ε2 allele vs. ε3 allele (P_{heterogeneity} = 0.054, I² = 44.70%) and ε2/ε2+ε2/ε3 vs. ε3/ε3 (P_{heterogeneity} = 0.071, I² = 39.50%), so Galbraith plot analysis and sensitivity analysis were performed to detect the possible sources of heterogeneity. Under the genetic model of ε2 allele vs. ε3 allele, the Galbraith plot analysis (Figure 4A) showed that the Halim et al. study was the outlier, which is consistent with the results of sensitivity analysis (Figure 4B). No heterogeneity existed after this outlier study was omitted (P_{heterogeneity} = 0.460, I² = 0%). Thus, the study by Halim et al. may be the source of heterogeneity in the meta-analysis for the ε2 allele vs. ε3 genetic model.

Similarly, under the genetic model of ε2/ε2+ε2/ε3 vs. ε3/ε3, the Galbraith plot analysis (Figure 4C) and sensitivity analysis (Figure 4D) indicated that Halim and Chaudhary's study were the outliers. When the two outlier studies were omitted, no heterogeneity existed in the remaining studies (P_{heterogeneity} = 0.681, I² = 0%). Therefore, the studies of Halim et al. and Chaudhary et al. may be the main contributors to the source of heterogeneity in the meta-analysis for the ε2/ε2+ε2/ε3 vs. ε3/ε3 genetic model.

No obvious asymmetry was observed in the shape of the funnel plot for the following genetic models: ε2/ε2 vs. ε3/ε3 (Figure 5A); ε2/ε3 vs. ε3/ε3 (Figure 5B); ε2/ε4 vs. ε3/ε3 (Figure 5C); ε4/ε4 vs. ε3/ε3 (Figure 5D); ε2 allele vs. ε3 allele (Figure 5E); ε2/ε2+ε2/ε3 vs. ε3/ε3 (Figure 5F). In addition, the Begg's test and Egger's test also did not show any evidence of publication bias (P_Begg = 0.251 and P_Egger = 0.08 for ε2/ε2 vs. ε3/ε3, P_Begg = 0.373 and P_Egger = 0.320 for ε2/ε3 vs. ε3/ε3, P_Begg = 0.283 and P_Egger = 0.403 for ε2/ε4 vs. ε3/ε3, P_Begg = 0.466 and P_Egger = 0.988 for ε4/ε4 vs. ε3/ε3, P_Begg = 0.119 and P_Egger = 0.053 for ε2 allele vs. ε3 allele, P_Begg = 0.300 and P_Egger = 0.331 for ε2/ε2+ε2/ε3 vs. ε3/ε3).

The results from the following three genetic models ε3/ε4 vs. ε3/ε3; ε3/ε4+ε4/ε4 vs. ε3/ε3 and ε4 allele vs. ε3 allele performed by Begg's test (P_Begg = 0.213, P_Begg = 0.033, and P_Begg = 0.043, respectively) or Egger's test (P_Egger = 0.013; P_Egger = 0.001 and P_Egger = 0.001, respectively) revealed publication bias. Nevertheless, by using the trim and fill method, the recalculated estimates (OR = 1.50, 95%CI = 1.24–1.82; OR = 1.59, 95%CI = 1.34–1.89 and OR = 1.40, 95%CI = 1.22–1.62, respectively) remained statistically significant, which indicated that our meta-analysis results were steady and not influenced by publication bias. Figure 6 shows the funnel plot of trim and fill method in the genetic model of ε3/ε4 vs. ε3/ε3 (Figure 6A), ε3/ε4+ε4/ε4 vs. ε3/ε3 (Figure 6B), ε4 allele vs. ε3 allele (Figure 6C).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.