The entire contents of the present meta-analysis followed the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-analysis statement (16). Literatures were collected based on the online PubMed (ncbi.nlm.nih.gov/pubmed) and EMBASE (embase.com/#search) databases up to June 30th 2016, utilizing the following search words: ‘Nasopharyngeal carcinoma,’ ‘microRNA,’ ‘miRNA’ and ‘diagnosis/sensitivity/specificity,’ with limitations to ‘human’ and ‘English.’ The inclusion criteria were as follows: i) Studies that evaluated the diagnostic performance of miRNA for NPC; ii) the final diagnosis of NPC was confirmed by tissue-proven histopathology; iii) studies provided complete data to construct 2×2 contingency tables; and iv) studies explicitly addressed the control sources and size. Studies with the following criteria were excluded: i) Studies without complete data to construct 2×2 contingency tables; ii) studies failed to explicitly state the control group(s); and iii) non-English written papers, review articles, basic study, letters, commentaries and meta-analyses.

Eligible data were retrieved by two authors, and the contents included the name of the first author, year of publication, origin, patient size, control sources and size, miRNA expression profiles, and sensitivity and specificity data. The included studies were further evaluated for the quality grading based on the diagnostic accuracy (Quality Assessment of Diagnostic Accuracy Studies; QUADAS) tool issued in 2003 (17). According to the QUADAS scoring criteria where 14 questions were included, a ‘Yes’ answer corresponded to a score of ‘1,’ whereas a ‘No’ or ‘Unclear’ answer received a score of ‘0.’

Statistical analyses were performed based on the platforms of Stata 12.0 (StataCorp LP, College Station, TX, USA) and Meta-disc 1.4 (Unit of Clinical Biostatistics Team of the Ramón y Cajal Hospital, Barcelona, Spain) software. Heterogeneity from the threshold effect was evaluated by the Spearman's correlation coefficient, and that from the non-threshold effect was assessed by Cochran's Q test and inconsistency index (I²) test (P<0.05 or I² >50%) (18). A random effects model was applied for the meta-analysis in the case of the existence of heterogeneity (19). The generated diagnostic parameters included pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR) and summary receiver operator characteristic (SROC) curve. The clinical utility (post-test probabilities) were assessed by Fagan's plot assays. Publication bias among studies was determined by Deeks' funnel plot asymmetry test with a significant level of P<0.05.

The process of study selection is demonstrated in Fig. 1. According to the predefined criteria, 275 relevant articles were obtained from online PubMed and EMBASE databases following elimination of the duplicated records. The retrieved papers received a detailed review of the study title and abstract, and 256 records were accordingly excluded. The possible eligible studies that received full text evaluation were restricted to 19, and 13 of them failed to meet the aim of the present study and were finally excluded. Eventually, six cohorts (including 21 individual studies) comprising 528 NPC cases and 252 matched controls were available for the meta-analysis (10–15).

The main features for each included study are summarized in Table I. All six studies were conducted in China, and the evaluation methods for miRNA levels were all based on reverse transcription-quantitative polymerase chain reaction. The test matrices comprised serum, plasma and tissue, and the reference genes contained U6, miR-454, miR-39 and miR-238.

The article quality of each included publication was evaluated in terms of the 14-item QUADAS checklist (17). All of the studies yielded a QUADAS score ≥8, suggesting a relatively high quality of the enrolled studies. The proportions of studies with low, high or unclear concerns are displayed in Fig. 2.

Table II summarizes the evaluated study heterogeneity from threshold and non-threshold effects. As indicated in Table II, the P-value obtained from Spearman's correlation coefficient in the overall pooled studies was estimated to be 0.012, indicating a significant heterogeneity generated from threshold effect. Additionally, the Cochran's Q test for the overall pooled analysis yielded a P-value of P<0.0001, along with an I² value of 83.8%, suggesting that the non-threshold effect is likely to be a source of heterogeneity as well. As a result, a random analysis model was selected for the final meta-analysis.

As indicated in Table III, the pooled sensitivity, specificity, PLR, NLR, DOR and area under the curve (AUC) for miRNA profiling were 0.76 [95% confidence interval (CI), 0.70–0.81], 0.76 (95% CI, 0.69–0.82), 2.81 (95% CI, 2.19–3.61), 0.35 (95% CI, 0.28–0.44), 9.01 (95% CI, 5.62–14.44) and 0.83, respectively. Following adjustment of the outlier studies, the combined sensitivity, specificity, PLR, NLR, DOR and AUC for miRNA profiling were estimated to be 0.78 (95% CI, 0.76–0.80), 0.78 (95% CI, 0.75–0.81), 3.19 (95% CI, 2.48–4.11), 0.32 (95% CI, 0.25–0.40), 11.43 (95% CI, 7.02–18.61) and 0.84, respectively. Forest plots of pooled sensitivity, specificity as well as the SROC curve are demonstrated in Fig. 3A-C.

For the clinical utility assessed by Fagan's plot assay, apparent improvements of post-test probabilities were displayed in the pooled analysis, with a post-test probability of a positive result of 45% and negative result of 7% (Fig. 3D).

Subgroup analysis was stratified by miRNA profiling (single or in parallel) and the test matrix. As exemplified in Table III, paralleled testing of miRNA achieved better diagnostic efficacy than single miRNA analysis: The pooled sensitivity was 0.88 (95% CI, 0.85–0.90) vs. 0.70 (95% CI, 0.68–0.72), specificity was 0.85 (95% CI, 0.81–0.89) vs. 0.69 (95% CI, 0.65–0.72) and AUC was 0.95 vs. 0.75. When the studies were stratified by the test matrix, the data manifested that serum/tissue-based tests conferred higher accuracy than plasma-based miRNA analysis in confirming NPC (sensitivity, 0.91 vs. 0.73; specificity, 0.94 vs. 0.71; AUC, 0.97 vs. 0.78).

Influence analysis and meta-regression tests were performed to deeply trace the heterogeneity sources. As demonstrated in Fig. 4, three individual studies were evaluated to undergo a deviation status. Following adjustment of the data by eliminating the outliers, the P-value of Spearman's correlation coefficient was altered from 0.012 to 0.021, and the I² value declined from 83.8 to 77.6%, hinting that the deviation data is likely a source of heterogeneity (Table II).

The meta-regression test was conducted based on five pre-specified covariates: Test matrix (plasma, serum or tissue), reference gene (U6 or other), test pattern (panel or single), NPC cases (<100 or ≥100), control size (<50 or ≥50) and article quality (QUADAS≥10 or <10) (20). The results revealed that the test pattern (P=0.0001) and reference gene (P=0.0026) were the key factors that contributed to the heterogeneity sources (Table IV).

The funnel plots for publication bias demonstrated no asymmetry for the overall pooled analysis, along with a P-value of 0.702, indicating that there was no bias from the publications (Fig. 5).