Retrieval data were extracted from papers published in English on PubMed, EMBASE, Web of Science and CNKI databases due November 1st, 2021. The key words encompassed “colorectal cancer”, “colorectal carcinoma”, “carcinoma of colon”, “colorectal neoplasms”, “circular RNA”, “circRNA”, “hsa circ”, “diagnoses”, “diagnosis”, “sensitivity”, “specificity”, “area under the curve”, “AUC”, “ROC curve”, “prognoses”, “prognosis”, “survival”, “overall survival”, “progression free survival”, “hazard ratio”, “OS”, “PFS” and “HR”. Besides, references attached to the paper were also searched manually to prevent omission of any eligible literature.

Inclusion criteria were listed as follows: 1) case-control studies; 2) studies on the evaluation of diagnostic value and/or prognosis of circRNA in CRC; 3) data of true positive number (TP), false positive number (FP), false negative number (FN) and true negative number (TN) that could be obtained directly or indirectly to construct a 2 × 2 four-grid table for the diagnostic meta-analysis; 4) the prognostic observation indice was total survival time (overall survival, OS), or progression free survival (PFS), for outputting the hazard ratio (HR) and 95% confidence interval (CI) directly or indirectly. Exclusion criteria were defined as follows: 1) a study size of less than 20 cases; 2) data used for statistical analysis were insufficient or unavailable even after contacting original authors; 3) non-English language articles or low quality research.

All relevant literature was screened by two trained authors. The basic information was extracted independently: name of the first author, date of publication, study population, sample size, the control type, circRNA signatures, detection methods, reference genes, cut-off value settings, follow-up time, as well as values of sensitivity, specificity, AUC, HR, and 95% CI(s) and so forth.

The Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 tool was adopted to evaluate the quality of references prior to the diagnostic meta-analysis (43). The evaluation system comprised two parts: bias evaluation and applicability. Specifically, the bias evaluation consisted of case selection, index tests, reference standards, and flow and timing, and the applicability evaluation of case selection, trials to be evaluated and gold standard. Each item could be classified as low risk, high risk and unknown, with the corresponding scores of 1, 0, and 0, respectively. The total score of ≥4 (with a full score of 7) indicated that the quality of literature research was high. The case-control study was evaluated according to 8 items of the Newcastle-Ottawa Scale (NOS scale) (44), categorized into study selection, comparability, and outcome. The total rated score of ≥5 (with a full score of 9) suggested that the quality of literature research was high.

Statistical analysis was carried out using MetaDiSc 1.4 and STATA 12.0 software, and the combined-effect indices involved sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), AUC, HR and 95% CIs. The difference in the merged area under the curve (AUC) between groups was compared by Z test. Spearman correlation coefficients were calculated to evaluate the threshold effect using MetaDiSc 1.4, while Cochran’ Q and I ² tests to assess the non-threshold effect using STATA 12.0. The level of significance was set at p < 0.01 or I ² > 50%. When there was no heterogeneity between studies, the statistics could be merged by the fixed-effect model; and when heterogeneity appeared, the statistics would be combined using the random-effect model. The sources of heterogeneity were deeply traced by a sensitivity analysis and a meta-regression test. Publication bias was judged by Deek’s quantitative funnel plot, as well as Begg’s and Egger’s tests, and the significant level was set at p < 0.1. The trim-and-fill method was adopted to assess the possible effect of publication bias (45).

Of the initially screened 536 references from the aforesaid online databases according to the retrieval strategy, 29 articles (19 for diagnosis and 14 for prognosis) (14–42) that met the inclusion and exclusion criteria were eventually enrolled for our meta-analysis. The literature retrieval process was depicted in Figure 1.

The basic characteristics of the 29 articles (14–42) were summarized in Tables 1, 2. A total of 2639 CRC patients and 1471 matched controls were enrolled. The CRC cases were confirmed by pathological examinations and the included control group encompassed healthy subjects and paracancerous controls. All tissue and plasma samples were obtained preoperatively without any other treatment. The included individuals in our study included Asians and Caucasians, and 34 circRNAs were involved in the meta-analysis, of which 25 acting as oncogenes were up-regulated and 9 as tumor-suppressor genes were down-regulated in CRC. The expression levels of circRNAs in CRC were determined by RT-qPCR or RNA sequencing, and GAPDH, 18srRNA, or β-actin mRNAs were used as internal reference genes. Of the 15 included studies on the prognosis of CRC, 9 provided available HRs and 95% CIs, 6 offered relevant data that could indirectly calculate the indices by formulas or prognosis curves. The median follow-up time varied from 1 to 39 months.

The Spearman correlation coefficient analysis showed a p value of 0.356 (Spearman correlation coefficient: −0.267) in diagnostic meta-analysis, suggesting that there was no heterogeneity resulted from the threshold effect. A p value of <0.001 and I ² of 97.46% for the overall combined diagnostic effect were presented in Cochran’ Q and I ² tests, indicating that substantial heterogeneity existed in the non-threshold effect. No heterogeneity was observed in the combined prognostic meta-analysis (oncogenic circRNAs: I ² = 0.0%, p = 1.000; adjusted effect of the tumor-suppressor circRNAs: I ² = 0.0%, p = 0.994).

For the diagnostic meta-analysis, the QUADAS-2 scale was used to assess the risk of bias. As a result, the rated scores of all 19 articles were higher than 4 (Table 3), suggesting that the overall quality of the included studies was high. Consistently, the NOS scale showed high scores of over 6 in the observational studies (Table 4), indicating the high quality of the included case-control studies.

The forest plots showed that the pooled sensitivity, specificity, PLR, NLR, DOR and AUC of circRNAs in the diagnosis of CRC were 0.75 (95% CI: 0.69–0.80), 0.74 (95% CI: 0.69–0.78), 2.87 (95% CI: 2.46–3.34), 0.34 (95% CI: 0.28–0.42), 8.37 (95% CI: 6.32–11.09) and 0.81, as shown in Figure 2. Subgroup analysis showed that the AUC of up-regulated circRNAs in the diagnosis of CRC was higher than that of down-regulated circRNAs (AUC: 0.83 versus 0.77; Z test, p < 0.05), and the former had higher diagnostic specificity and DOR (Table 5). Analysis based on different control sources showed that the efficacy of circRNA expression profile was higher in distinguishing CRC from healthy differentiation than its ability to distinguish CRC from PNC (Table 5). Analysis based on different sample type and reference gene showed that the efficacy of circRNA abnormal expression profile based on issue was equivalent to that of ACU based on plasma/serum; GAPDH based testing and non-GAPDH based testing could also obtain similar results (Table 5).

The subgroup classification was conducted according to the gene functions of circRNAs. The prognosis analysis showed that high expression levels ofup-regulated circRNAs were associated with the poor OS in GC patients (HR = 2.38, 95% CI: 1.66–3.41, p = 0.000), whereas the total survival time in GC patients with high expressed down-regulated circRNAs was significantly prolonged (an outlier eliminated adjusted HR = 0.33, 95% CI: 0.15–0.72, p = 0.006) (Figure 3).

Sources of heterogeneity were traced by a sensitivity analysis, and no outliers were identified in the diagnostic meta-analyses and the prognostic effect for the plasma/serum-based circRNA testing (Figures 4A–C). However, it was found that data of one independent study of the prognostic effect was identified as an outlier (18) (Figure 4D). After the removal of the outlier, the p value of Cochran’ Q test varied from 0.000 to 0.961, and I ² decreased to 0.0% from 29.0%. All this suggested that the inclusion of outliers for statistical consolidation could be considered as one of the important reasons for the heterogeneity of the analysis results. Thus, the adjusted HR of the down-regulated circRNAs was calculated after an elimination of the outlier (Figure 3C).

On the other hand, a meta-regression test was performed to analyze factors comprising the quality of the study, the type of study, the type of specimen, the number of cases and the number of control groups. The results showed that these factors were not the source of heterogeneity among the studies with nonsignificant differences (data not shown).

Assessment of the publication bias is shown in Figure 5. We observed publication bias in the combined prognostic effect forup-regulated circRNAs (Egger’s test, p = 0.021) (Figure 5C). The missing studies were adjusted through the trim-and-fill method (Figure 5D). Nevertheless, the pooled analysis incorporating the hypothetical studies (moment-based estimate of between studies variance = 0.000, p = 1.000) altered slightly from the unadjusted ones (moment-based estimate of between studies variance = 0.000, p = 1.000), suggesting that the bias did not have a significant impact on the combined effect ofup-regulated circRNAs in CRC.