Systematic review and meta-analysis of tumor biomarkers in predicting prognosis in esophageal cancer

To identify all primary research studies that evaluated levels of candidate biomarker expression as a prognostic factor among individuals with EC, we searched the PubMed medical literature database up to April 11, 2012, without language restrictions, using a strategy developed with an expert librarian based on terms for esophageal carcinoma, prognostic studies [10] and biomarkers. The search strategy was based on combinations of (“esophageal” or “esophagus”), (“neoplasms” or “carcinoma” or “tumor”), (gene or protein or biomarker or marker), and (prognos* or “survival analysis” [Mesh] or “follow-up studies” [Mesh] or mortality [Majr] or mortality[subheading] or incidence [Mesh] or predict or course or outcome). One reviewer (M. Chen) obtained the full texts of relevant articles following the search and inspection of titles and abstracts of citations to identify those articles that were likely to report the study of prognostic biomarkers in EC. In cases where data in several publications were derived from part or all of the same patient series, only the study presenting the most recent or most complete dataset was included.

We used published guidelines for reporting tumor marker studies and quality metrics for evaluating studies to include in the cancer-related meta-analyses [9, 11]. Criteria used to determine study eligibility were as follows: 1) a prospective or retrospective cohort design with a well-defined study population and justification for all excluded eligible cases, 2) assay of the primary EC specimens, 3) a clear description of methods for specimen handling and testing, including selection and preparation of reagents or kits, as well as visualization techniques, 4) clear statements on the choice of positive/present and negative/absent controls and on assay validation, 5) statistical analysis using multivariable proportional hazards modeling that adjusted for clinical prognostic factors, and 6) reporting of the resultant adjusted hazard ratios (HRs) and their 95% confidence intervals (CIs), or provision of data available for statistical estimation of HRs. Because esophageal small cell carcinomas, epidermoid cancer of the esophagus, and neuroendocrine carcinoma of the esophagus have different clinical courses, studies that did not distinguish these tumor types from EC were excluded.

Quality assessment was performed in duplicated for each eligible study by two independent reviewers (Chen and Huang) using operationalized prognostic biomarker reporting guidelines [9] and extract details on 16 items (Additional file 1: Table S1). This scale allowed for assessment of study design, biomarker measurement, outcome and analysis.

Two investigators (Chen and Zhu) reviewed all eligible studies and carefully extracted study characteristics in duplicate, including the first author’s name, publication year, country of origin, histology, sample size, gender, mean/median age, disease stage, test method, cutoff value, the status of biomarker expression, and the computed multivariable hazard ratio and its 95% CI. When results were present without confidence intervals, the p value was used to estimate the confidence intervals via the z-statistic.

All eligible individual biomarker assays were sorted according to their major biological function. Function was determined by reviewing the current scientific literature comprehensively and then classifying according to the 5 acquired capabilities of cancer as defined by Hanahan and Weinberg [12], and which included sustained angiogenesis, evasion of apoptosis, insensitivity to antigrowth signals, limitless replicative potential and tissue invasion and metastasis. To accommodate blood biomarkers, the Hanahan-Weinberg classification system was supplemented by one additional category: serum markers. Biomarkers evaluated in less than five studies, were excluded in this review and out of quantitative synthesis. For biomarkers assayed in five or more studies, the summary HR and 95% CI were calculated by using fixed effects according to generic inverse variance and random effects model using the DerSimonian-Laird method [13]. Statistical heterogeneity among studies was assessed using Q and I² statistics [14]. We considered that heterogeneity was present when the Q-test P-value was less than 0.1. In addition, when I² was lower than 50%, studies with an acceptable heterogeneity were considered, and the fixed-effects model was used; otherwise, a random effect model was adopted. The combined HRs were estimated graphically by Forest plots. Possible source of heterogeneity were investigated by subgroup analysis. Study publication bias was assessed with counter-enhanced funnel plots, by Begg’s adjusted rank correlation test and by Egger’s regression asymmetry test [15‐17]. When p > 0.05 was considered to indicate that there was no publication bias in the studies. All statistical analysis was conducted using Stata SE 11.0 software (Stata Corporation).

The abstracts and titles of 3259 primary manuscripts were identified for initial review using strategies as described. Reviewers identified 979 manuscripts to be appropriate in terms of evaluation of prognostic biomarkers in EC. For these manuscripts, full-text articles were obtained. Upon further review, 109 studies published between 1994 and 2012 were eligible for this systematic review and with meta-analysis (Figure 1).

All reported the prognostic value of biomarkers in patients with EC by presenting multivariable survival estimates for differential levels of candidate biomarker expression. Effective sample size ranged from 29 [18] to 708 patients [19] (median, 87 patients), with 13 studies including 50 or fewer patients, 54 studies including 51-100 individuals, 26 studies including 101-150 individuals, 12 studies including 151-300 individuals and 4 studies including more than 300 individuals. Seventeen clinicopathologic factors were incorporated in one or more of the eligible studies’ multivariate analysis. The most commonly included prognostic factor was depth of invasion involvement with lymph node status being included in 67 (61%) studies and 65 (59%) studies. Other common adjustment parameters included tumor stage (37 of 109 studies), gender (27 of 109 studies) and metastatic status (27 of 109 studies) (Figure 2A). Fifty-seven studies considered three to five clinical parameters in their multivariable proportional hazards models, 26 studies considered less than three covariates, 21 studies included more than five covariates and another 5 studies did not report(Figure 2B).

These 109 studies presented data on 13 unique biomarkers. The majority of eligible studies (n = 87) restricted their analysis to a single included candidate marker, and the remaining 22 evaluated between two to five markers. Regarding angiogenesis, 6 studies of cyclooxygenase-2 (COX-2) [20‐25] and 19 studies of vascular endothelial growth factor (VEGF) [26‐44] were included. Of the eligible marker associated with apoptosis, 5 studies of survivin [45‐49] were available for analysis. For cell-cycle regulators, 9 studies of p21 [27, 50‐57] and 7 studies of p27 [25, 58‐63] were included. Four markers associated with replicative potential were eligible for examination, and were comprised of 15 studies of cyclin D1 [27, 51, 54, 58, 59, 62, 64‐72], 32 eligible studies of p53 [27, 30, 32, 34, 38, 40, 51, 54],[68, 73‐95], 6 studies of human epidermal growth factor receptor-2 (HER-2) [19, 96‐100] and 5 studies of Ki67 [25, 27, 87, 101, 102]. Regarding the tissue invasion and metastasis markers evaluated, 10 studies of E-cadherin [27, 65, 102‐109] were included. Three eligible serum markers were eligible for analysis, and were comprised of 8 studies of C reative protein (CRP) [110‐117], 5 studies of SCC-Ag [18, 118‐121] and 5 studies of hemoglobin (Hb) [122‐126]. The 13 biomarkers were evaluated for overall survival (OS) and sorted according to five to eight Hanahan-Weinberg functional capabilities modified to include serum markers (Additional file 2: Table S2). Additional file 2: Table S2 outlines the demographic, clinicopathological, methodological and outcome characteristics of these studies.

The mean number of study quality items reported was 11 out of possible 16 and was not associated with sample size with correlation coefficient of 0.06, p = 0.52 (Figure 3). There was also no statistically significant difference between the quality items of 65 positive studies (where a biomarker was statistically a factor for poor prognosis) and 44 negative studies (where a biomarker was not statistically significant) (Mann-Whitney p = 0.27). All of the studies reported details of the clinical endpoint and multivariate analysis. More than 90% of studies reported details of the objective or prespecified hypothesis, patient source, population characteristics, assay method, manufacturer, cutpoint and confounders. Of note, 63 studies reported the follow-up period or the median follow-up time. One study referred to a missing value, but no study referred to a statistical sample size. To try to evaluate the impact of study quality on the final pooled estimate, a subgroup analysis was performed according to the different number of quality items: eleven or more eleven, which is the mean number of study quality items, and less than eleven. For majority markers, the results were consistent, the pooled HRs were not significantly altered, suggesting the study quality improbable as source of bias. For COX-2, only one study with quality items less than eleven, the HR was reported 2.34 (95% CI, 1.11-4.91). The pooled HR estimated for the other five studies was 1.36 (95% CI, 0.60-3.08), which is similar to the final pooled HR 1.54 (95% CI, 0.80-2.98). For HER-2, both subgroups included three studies eligible for meta-analysis, with combined HR 1.64 (95%CI,1.07-2.51) and 1.26 (95% CI, 0.70-2.26), respectively. Because of a small number of studies included, the results should be treated with caution.

For the 13 markers, a multivariate HR and 95% CI were available from five or more studies and were combined using both fixed effects general inverse variance and DerSimonian-Laird random effects modeling to obtain a single summary HR and 95% CI. (Table 1) Five of the six original Hanahan-Weinberg functional capabilities along with the additional group were represented by at least one marker statistically associated with OS. Two studies [35, 91] were excluded which failed to present validated data in their meta-analysis.

Of the six studies that used immunohistochemistry (IHC) for OS data, the pooled HR was 1.54 (95% CI, 0.80-2.98), with significant evidence of heterogeneity between the contributing studies (I² = 75%). Restricting analysis to the four studies assessing COX-2 expression in ESCC, the pooled HR was 0.96 (95% CI,0.39-2.41), again, with evidence of study heterogeneity (73.5%). These results should, however, be interpreted with caution because of the small number of contributing studies and the significant evidence for significant study heterogeneity. Two studies assessed COX-2 expression in EADC displayed a pooled HR of 3.06 (95% CI, 2.01-4.65), with no evidence of heterogeneity (I² = 0.0).

Of the eighteen VEGF expression studies eligible for pooling of OS data, the pooled HR was 1.80 (95% CI, 1.51-2.14) with no evidence of heterogeneity. The Forrest plot for this analysis is shown in Figure 4A. When restricting analysis to the sixteen studies examining VEGF expression in ESCC, the combined HR was 1.85 (95% CI, 1.55-2.21) with no evidence of heterogeneity. Of the other two studies, one presented data on EC and one on EADC, respectively. To assess the effect of four methods on evaluating VEGF expression, we pooled HRs from studies using IHC, enzyme-linked immunosorbent assay (ELISA) or reverse transcription polymerase chain reaction (RTPCR). ELISA-based studies demonstrated a larger pooled HR (HR 2.55, 95% CI, 1.72-3.79) compared to IHC-based studies (HR 1.59, 95% CI, 1.28-1.98) or RTPCR-based (HR 1.90, 95% CI, 1.24-2.91) studies.

For five studies evaluating survivin expression in esophageal cancers, the combined HR was 1.90 (95% CI, 1.06-3.40) and there was evidence for heterogeneity (I² = 74.6%). The pooled HR estimated for survival in the four IHC-based studies involving ESCC was 1.57 (95% CI, 0.91-2.69), again, with significant evidence of heterogeneity (I² = 70.7%). The remaining PCR-based study involved EC, and had an HR of 6.60 (95% CI, 1.97-22.12).

Nine studies examined p21 and seven studies assessed p27 levels with the pooled HRs of 1.27 (95% CI, 0.75-2.16) and 1.68 (95% CI, 0.90-3.12), respectively, and there was evidence of heterogeneity within both groups. All studies used IHC to estimate the correlation between biomarker expression and survival. Interestingly, when grouped according to the histology of individual studies, the combined HR in ESCC for p21 (seven studies), remained unchanged at 1.28 (95% CI, 0.70-2.33). The pooled HR for p27 (six studies) in ESCC was 1.97 (95% CI, 1.00-3.88). In both analyses, there was significant evidence of study heterogeneity (I² = 87.4% for p21 and I² = 74.6% for p27, respectively).

Fifteen studies assessed cyclin D1. The overall pooled HR was 1.73 (95% CI, 1.34-2.23) and there appeared to be some heterogeneity between the studies (I² = 56.3%) Figure 4B illustrates the Forrest plot for the pooled data. In subgroup analysis, the pooled HR for ESCC was 1.82 (95% CI, 1.50-2.20) with no evidence of heterogeneity. Only three of the fifteen studies presented data evaluable for assessment of EC and the pooled HR was 1.18 (95% CI, 0.57-2.45). However, this result should be interpreted with caution because of the small number of contributing studies and evidence for significant study heterogeneity (I² = 77.4%). To assess the effect of the method used to assess cyclin D1 expression, HRs, using either IHC or PCR, were pooled. This presented a large pooled HR for PCR-based studies (HR 2.62, 95% CI, 1.41-4.89) compared to that from the IHC-based studies (HR, 1.64, 95% CI, 1.26-2.14). The IHC-based group displayed significant heterogeneity, whereas the PCR-based group did not.

Six studies examined the HER-2 as a biomarker. When conducting subgroup analysis, we found three of six eligible studies assessed HER-2 expression in EADC and had a pooled HR of 2.15 (95% CI, 1.39-3.33) with no evidence of heterogeneity. Two studies assessed HER-2 expression in an EC setting with a pooled HR of 0.91 (95% CI, 0.73-1.12). Another study examined HER-2 in ESCC and reported an HR of 0.92 (95% CI, 0.35-2.41). Because of the small number of studies included, the results should be treated with caution.

Three of five studies eligible for assessing Ki-67 in ESCC demonstrated a pooled HR of 1.11 (95% CI, 0.70-1.78) with no significant of heterogeneity. None of the HR reported were statistically significant except in one study that assessed in EADC and reported an HR of 0.26 (95% CI, 0.11-0.60).

Thirty-one studies assessed p53. The pooled HR of 1.34 (95% CI, 1.21-1.48) revealed significant association with overall survival and the Forest plot for this analysis is shown in Figure 4C. Restricting analysis to the twenty studies assessing p53 expression in ESCC gave a pooled HR of 1.26 (95% CI, 1.11-1.42), with ELISA-based studies giving a larger pooled HR (2.13, 95% CI, 1.46-3.09) than IHC-based studies (HR, 1.25, 95% CI, 1.12-1.40).

Ten studies assessing the E-cadherin biomarker displayed a pooled HR of 1.13 (95% CI, 1.06-1.21) with evidence of heterogeneity (Forrest plot in Figure 4D). When restricting analysis to the seven studies examing E-cadherin in ESCC, however, the result remained unchanged, with an HR of 1.12 (95% CI, 1.05-1.19), again, with significant heterogeneity. For the two studies assessing E-cadherin expression in EC, the pooled HR was 1.41 (95% CI, 1.05-1.89), and there appeared to be no heterogeneity between the studies. An EADC study reported an HR of 3.30 (95% CI, 0.99-10.99). When grouped according to method, the combined HR of IHC-based studies was 1.28 (95% CI, 1.11-1.49) with significant heterogeneity. The only ELISA-based study reported an HR of 1.10 (95% CI, 1.02-1.18).

Five eligible studies assessed SCC-Ag expression by enzyme immunoassay (EIA) in ESCC, and the pooled HR for OS was 1.28 (95% CI, 0.97-1.69) with no evidence of heterogeneity. Eight studies examined CRP expression with pooled HR 2.65 (95% CI, 1.64-4.27) and there was a significant heterogeneity. When conducted subgroup analysis, both EC and ESCC group still showed evidence of heterogeneity. When grouped according to different method used to assess CRP expression, two IHC-based studies had pooled HR 4.33 (95% CI, 2.02-9.24). Other studies used different method reported HR revealing a significant association with poor survival but one ELISA-based study (HR, 1.42, 95% CI, 0.83-2.42).

Five studies assessed Hb. The pooled HR was 0.91 (95% CI, 0.83-1.00) with significant heterogeneity. When restricting analysis to the three studies assessing Hb levels in EC, the pooled HR was 0.96 (95% CI, 0.89-1.03), again, with evidence of heterogeneity. Two studies assessing Hb levels in ESCC gave a pooled HR of 0.54 (95% CI, 0.40-0.74) with no evidence of heterogeneity. Only one included study reported no significant association with outcome. However, data were not sufficient to determine the prognostic value of Hb expression in either ESCC or EADC.

We performed sensitivity analyses, in which one study was removed at a time, to evaluate the result stability. For COX-2, VEGF, cyclin D1, p53, E-cadherin and SCC-Ag, the results indicated that fixed-effects estimates and/or random effects estimate before and after the deletion of each study were similar at large, suggesting high stability of the meta-analysis results. For survivin and CRP, although the results are consistent with the overall pooled estimates, the influencing single study conducted by S. Mega et al. and Ines Gockel et al. respectively. For other markers, the sensitivity analysis did not indicate high stability of the results due to one or two studies. The results of sensitivity analyses are the supplement of the subgroup analysis results.

Begg’s test and Egger’s test were used to examine publication bias. There was evidence for significant publication bias with p21 (Egger’s test p = 0.040), HER-2 (Egger’s test p = 0.042) and CRP (Egger’s test p = 0.005).