We performed an online search for relevant studies using the following databases: PubMed, China National Knowledge Infrastructure, Web of Science, Embase, Wanfang, and Google scholar. The search date is updated to November 15th, 2019. We placed language restrictions in Chines and English. The following keywords were employed for literature retrieval: (“nasopharyngeal carcinoma” OR “NPC” OR “nasopharynx cancer”; “EB virus OR EBV”, “ZEBRA OR Zta,” “diagnoses” OR “diagnostic value” OR “sensitivity and specificity” OR “ROC curve” OR “receiver” operating characteristics. We also reviewed the references lists of relevant studies for potential eligible studies. The latest data was used for the replicated data. We performed this meta-analysis by following the Preferred Reporting Items for Systematic review and Meta-Analysis checklist (Supplementary Material 1). The search strategy was presented in the Supplementary Material 2.

The included must meet the following criteria: (1) studies assessing the diagnostic of serum EB virus Zta antibody in the NPC; (2) studies with sufficient data for analysis, including true positive (TP), false positive (FP), false negative (FN), and true negative (TN); (3) NPC was confirmed by pathology examination; (4) for the same data, the latest results were used. The following studies were excluded: (1) studies without sufficient data were excluded. (2) Study was settled in a specific population. (3) Reviews, letters, comments, editorials, and case reports were also excluded.

Two investigators interpedently extracted the data and the disagreements were solved by consensus. For each study, the following data was extracted: the surname of the first author, publication year, ethnicity, sample size including case and control, examination methods, gold standard, 4-folds data (TP, FP, FN, TN), and sensitivity and specificity of each study.

We used the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) to assess the quality of included study (13). This tool has been widely used in the diagnostic accuracy studies. The QUADAS tool assessed the study quality from two domains: risk of bias and concerns about application. The risk of bias consists of four sub-items: patient selection, index test, reference standard, and flow of patients through the study and timing of the index tests and reference standard. Risk of bias is considered as “high” “unclear” or “low.” If the answers to all sub-items for a domain are “yes,” then risk of bias can be treated as low. If any of sub-item is judged “no,” potential bias may exist. Concerns about applicability are scored as “high” “unclear” or “low.”

First, we assessed the threshold effect using the Pearson coefficient (r = 0–0.341, P = 0.103). The threshold effect determined which model was used (14). No threshold effect existed for the present study. And the bivariate mixed effects model was used. We calculated the following parameters and their 95% confidence internals (CIs): sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnosis odds ratio (DOR), and summary receiver operating characteristics curve (AUC), An AUC of 1.0 represents the perfect discrimination ability (15–17). The heterogeneity within studies was examined using Q test and I² statistic. P < 0.05 and I² > 50% indicated the significant heterogeneity (18, 19). Fagan' plot and the line graph of post-test probabilities vs. prior probabilities between 0 and 1 using summary likelihood ratios (20). Sensitivity analysis: quantile plot of residual-based goodness-of fit and Chi-squared probability plot of squared Mahala Nobis distances were used for assessment of the bivariate normality assumption; spike plot was used for checking for particularly influential observations using Cook's distance. Scatterplot was used for checking for outliers using standardized predicted random effects. The publication bias was assessed by Deek's funnel plot asymmetry test (21). No overlap between two confidence intervals indicated significant difference. All analyses were completed on Stata 14.0 and Reviewer manager 5.0. P < 0.05 was considered as significant level.

We totally obtained 358 articles from six online electronic database. 110 articles were excluded because of duplicates data and publications. We checked the titles and abstracts of 248 articles and removed 196 articles because they are significantly unrelated topics and others publications, such as reviews and comments. We downloaded the full-text of 52 articles for further screening. Among of these articles, seven studies with insufficient data, three articles with unrelated topics or diagnostic values, and nine articles belonged to reviews, comments, letter and meeting abstract. At last, we included 23 studies including 24 pieces of data (Supplementary Material 3). The selection flow of study selection is presented in Figure 1. The total sample size is 17,770 with 2,126 cases and 15,644 controls. These studies were published from 2003 to 2018. All cases were confirmed by pathology examination. The examination of antibody was ELISA. The highest sensitivity was 0.96 and the lowest was 0.36. The highest specificity was 0.97 and the lowest was 0.81. The distributions of 4-folds (TP, FP, TN, FN) and details were shown in Table 1.

The quality assessment of included studies was summarized in the Figures 2A,B. Figure 2A presented the risk of bias and applicability concerns of each study. One study was judged as high-risk bias because of the flow and timing issue. One study was considered as high-risk bias because of the index test. The ratio of studies with high-risk bias was 8.7%. Two studies were judged as unclear risk bias in the item of flow and timing. Two studies had unclear risk bias in the item of index test. One study was scored as unclear applicability concern in the index test. The ratio of unclear risk studies was 17.4%. The ratio of studies with unclear concerns regarding applicability was 4.3%. On the whole, the included studies achieved a high point in the quality assessment.

Figures 3A,B presented the combined results of sensitivity and specificity. The overall combined sensitivity was 0.85 (95%CI: 0.80–0.89) and the combined specificity was 0.90 (95%CI: 0.87–0.93) that were from the random-effect model. Because the heterogeneity was high (P < 0.05 and I² > 50%). The summarized AUC was 0.94 with 95%CI of 0.92–0.96 (Figure 4). The PLR was 8.9 (95%CI: 6.4–12.2) and the NLR was 0.17 (95%CI: 0.12–0.23). The diagnostic odds ratio was 53 (95%CI: 32–87). According to the criteria, PLR > 10 and NLR < 0.1 indicated high accuracy. According the diagnostic criteria, the EBV Zta antibody examination achieved a high diagnostic ability for NPC. The Figure 5 shows the pre-test probability and post-test probability. Based on the PLR, the post-test probability could arrival at 69%.

Because the heterogeneity within studies is high. We also performed subgroup analysis in the publication year (Median ≤2006 vs. >2006), sample size (≤Median vs. >Median) and ethnic (Asian vs. others). For publication year, the sensitivity was 0.88 (95%CI: 0.84–0.91) and the specificity was 0.90 (95%CI: 0.84–0.93) for ≤2006. The AUC, PLR, NLR and DOR were 0.94, 8.8, 0.13, and 64. The pooled results were similar for >2006 group. The values of six parameters were 0.80, 0.91,0.93, 8.8, 0.22, and 40. For different sample size, the pooled AUC was 0.94 for ≤Median and was 0.95 for >Median that were very close to the overall estimations. For different population setting, no overlap was found in the sensitivity (0.84 vs. 0.87), specificity (0.90 vs. 0.84), PLR (8.7 vs. 5.5), NLR (0.16 vs. 0.08–0.33), DOR (49 vs. 35), and AUC (0.94 vs. 0.92) between Asian and others. The details of results were presented in the Table 2.

The sensitivity analysis was presented the Figure 6. The Goodness of fit plot (Figure 6A) and bivariate normality plot (Figure 6B) indicted studies distributed along the references line. The influence analysis (Figure 6C) suggested that only one study was beyond the reference line. The outlier detection (Figure 6D) indicated the same study the same as the influence analysis. Two study were out of the detection and the further pooled results found no significant change. All signs indicated the pooled results are stable and few studies can affect the pooling results. The Deeks's funnel plot (Figure 7) asymmetry test indicated no significance (P = 0.770), which means no publication bias exits.

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.