Figure 1 presents the flowchart of the literature search procedure. Searching PubMed, EMBASE, EBSCO, Biomed central, and CNKI databases, as well as other sources, resulted in an initial inclusion of 236 records after duplicates were removed. Two writers separately reviewed the titles and abstracts of 236 publications, and 187 records were excluded because their study contents were unrelated to cirRNAs in HCC. The remaining articles were intensively evaluated for the full-text contents, and 28 of them were evaluated as review articles, or basic studies with irrelevant data, or relevant articles with insufficient information, which therefore were all eliminated. In the final stage, only 21 studies, including 8 publications for diagnosis [11–13, 19–21, 25, 29], 11 for prognosis [10, 11, 14–18, 22, 26, 28, 30], and 14 for clinicopathological feature [11–13, 15, 17, 18, 22–24, 26–30], were included in the quality assessment and quantitative synthesis.
Characteristics of Included Studies
Study characteristics are shown in Tables 1 and 2. All of the included studies were identified as case-control studies, in which 8 studies with 712 HCC cases and 788 controls assessed the diagnostic performance of circRNAs in HCC, and 11 studies included 2719 cohorts focused on the evaluation of the prognostic value of cirNRAs. All HCC cases were reliably diagnosed based on histopathological methods. The control sources included chronic hepatitis, liver cirrhosis, para-tumorous tissues [11, 20, 21, 25], and non-cancer/healthy individuals [19, 29]. Amplification of circRNAs was enabled by using the qPCR test, and GAPDH or β-action were used for normalization. The circRNA signatures for diagnosis included hsa_circ_0003570, circZKSCAN1, hsa_circ_0005075, Hsa_circ_0001649, hsa_circ_0091582, hsa_circ_0128298, hsa_circ_0004018, hsa_circ_0001445, and circRNAs panel sets. CircRNA profiles for prognosis contained hsa_circ_0001649, circ-ITCH, circMTO1, cSMARCA5, circC3P1, hsa_circRNA_100338, hsa_circ_0064428, circRNA101368], hsa_circ_0103809, and circ-ZEB1.33.
Methodological Quality Assessment
The quality and bias of all diagnostic studies were independently appraised by two authors in compliance with the QUADAS-II criteria, whereby studies were assessed for patient selection, index test, reference standard, flow and timing . As reported by Figure 2, all included 8 publications for diagnosis were judged as low risk for applicability concerns, and 3 studies were assessed with bias in patient selection, or index test, or reference standard, and received rated QUADAS scores equal to 3 points. Evaluation of the quality of all case-control studies was enabled by applying the NOS checklist . As shown in Table 3, all the included prognostic studies received rated NOS scores higher or equal to 6, and thus they were all included in the final synthesis.
Investigations of Heterogeneity
In the overall diagnostic meta-analysis, the Chi2 and I2 tests revealed significant substantial heterogeneity among pooled effects (Q = 49.403, df = 2.00, p = 0.000; I2 = 95.95, 95% CI: 92.85–99.05). In line with the diagnostic effects, clear heterogeneity was also observed in the pooled prognostic effects for both the elevated (p = 0.000; I2 = 91%) and down-regulated circRNA profiles (p = 0.000; I2 = 92.7%). Thus, all weights were synthesized using a random effect model.
Overall Diagnostic Performance
The summary receiver-operating characteristic (SROC) curve was employed to assess the diagnostic efficacy of circRNA profiling in distinguishing HCC from non-tumorous controls. The pooled sensitivity (Figure 3A), specificity (Figure 3B), PLR (Positive Likelihood Ratio), NLR (Negative Likelihood Ratio), DOR (Diagnostic Odds Ratio) (Figure 3C) and AUC (Figure 3D) were estimated to be 0.78 (95% CI: 0.69–0.85), 0.80 (95% CI: 0.74–0.86), 3.97 (95% CI: 2.85–5.54), 0.27 (95% CI: 0.19–0.39), 14.59 (95% CI: 7.83–27.21), and 0.86, respectively.
We found distinct prognostic value in the abnormally expressed circRNA signature in HCC, wherein the signature covered up-regulated circRNAs and was negatively correlated with the OS of patients with HCC (HR = 2.22, 95% CI: 1.50–3.30, p = 0.000) (Figure 4A), hinting that these circRNAs could be considered as independent prognostic biomarkers in HCC. Meanwhile, the significantly higher survival time (OS) was found in HCC patients with down-regulated circRNA profiling (HR = 0.42, 95% CI: 0.19–0.91, p = 0.028) (Figure 4B), suggesting that circRNAs with decreased expression status were more prone to act as tumor suppressor genes in HCC.
Evaluation of the link between circRNA expression and clinicopathological elements in HCC also produced robust results. As shown in Table 4, significant associations were observed between the circRNA expression and alcoholism (pooled p = 0.0323), tumor size (pooled p = 0.00012), differentiation grade (pooled p = 0.000), microvascular invasion (pooled p = 0.003744), TNM stage (pooled p = 0.000), metastasis (pooled p = 0.000), and serum AFP level (pooled p = 0.0115).
The stratified analysis depended on sample type and revealed that the tissue-based circRNA testing expressed a slightly higher diagnostic efficacy in confirming HCC than the overall pooled analysis (AUC: 0.88 vs. 0.86; DOR: 15.17 vs. 15.59). Different effects were also observed in different expressed circRNAs, wherein up-regulated circRNA profiling yielded a better diagnostic performance than down-regulated circRNAs (AUC: 0.97vs. 0.81; DOR: 11.48 vs. 8.75). Moreover, the analysis grouped by control type expressed that circRNA profiling could differentiate chronic hepatitis or cirrhosis from HCC, with an AUC of 0.84, sensitivity of 0.77, and specificity of 0.76; additionally, the circRNA expression signature was able to distinguish adjacent non-cancerous liver tissues from HCC samples, with an AUC of 0.73 and specificity of 0.75 (Table 5).
Influence analysis was performed in both the diagnostic and prognostic effect sizes. As exemplified by Figure 5, one study  was identified as the outlier in the pooled prognostic effects of down-regulated circRNAs in HCC. After elimination of the outlier data and re-analysis of the effect, the I2 dropped from 92.3% to 90%, indicating that included heterogeneous studies were a substantial cause of study heterogeneity. No outliers were detected in other pooled effects (Figure 5).
Publication bias was judged using different methods for different pooled effects. As shown in Figure 6A, no clear publication bias was detected in the combined diagnostic effects (Deek’s funnel plot, p = 0.446), nor in the analysis of down-regulated circRNA profiling (Egger’s test, p = 0.606, Figure 6B). Nevertheless, the funnel plot expressed evidence of a publication bias in the effects of up-regulated circRNA profiling (Egger’s test, p = 0.001, Figure 6C), and the trim and fill technique was employed to detect the outcome of bias . As indicated in Figure 6D, the filled funnel plots identified 5 imputed studies, but the effect was slightly altered before and after adjustment (variance = 0.187, p = 0.005 vs. variance = 0.287, p = 0.000).