PCNSL is an aggressive tumor with a life expectancy of 3–5 months without treatment. Unfortunately, the definite diagnosis of PCNSL is often delayed by averagely 3 months after the initial symptoms appear for non-acquired immunodeficiency syndrome patients , because diagnosing PCNSL remains a challenge. Clinical and radiological features may suggest the suspected diagnosis of PCNSL, but they are not definitely diagnostic. Diagnosis of PCNSL is usually established by stereotactic brain biopsy. However, this invasive procedure has a complication rate of 8.5%, including haematomas, seizures or brain oedema . Although the cytology or flow cytometry of CSF or vitreous fluid is less invasive than stereotactic brain biopsy , these are only positive in case of leptomeningeal or ocular involvement respectively. In search of a novel diagnostic approach with high accuracy and limited risks to shorten the delay of PCNSL diagnosis, many researchers have turned their eyes to miRNAs, which can stably present in several body fluids including CSF and blood, and may serve as non-invasive biomarkers for PCNSL diagnosis [13, 26].
However, the researchers have found inconsistent results. Baraniskin et al.  reported the analysis of miR-21 (95.7% sensitivity and 83.3% specificity), miR-19b (95.7% sensitivity and 83.7% specificity), and miR-92a (95.7% sensitivity and 80.0% specificity) in CSF, either alone or combined (95.7% sensitivity and 96.7% specificity), identified PCNSL from controls with high accuracy. Nonetheless, Zajdel et al.  performed a combined assessment of the three miRNAs in discriminating PCNSL from non-malignant brain lesions, and found a specificity of 80.8% and a sensitivity of 63.3%, indicating lower diagnostic accuracy than Baraniskin et al. presented. Mao et al.  examined miR-21 in the blood and showed a sensitivity of 86.0% and a specificity of 90.0% in the detection of PCNSL. Yang et al.  found higher diagnostic accuracy of miR-21 in blood, with a specificity of 91.7% and a sensitivity of 96.3%. Therefore, we found it is necessary to assess the potential applicability and reliability of miRNAs as diagnostic biomarkers for PCNSL patients.
To the best of our knowledge, this is the first study to specifically evaluate the value of miRNAs as diagnostic biomarkers in PCNSL. Although there was one meta-analysis study is about miRNAs as diagnostic biomarkers in CNS cancers has been published , it focused on all types of CNS cancers, including but not limited to PCNSL. In addition, few studies of PCNSL were included in that meta-analysis study. It could not support the further evaluation on relationship between miRNAs and PCNSL consequently, which needs more PCNSL cases to confirm the final findings. Another relevant systematic review  did not perform a formal meta-analysis on the diagnostic accuracy of the markers for CNS lymphoma in blood and CSF, and focused on various markers at the same time including CXCL13, interleukins-6, -8, and − 10, soluble CD19 and so on, not just miRNAs. All in all, the specific evaluation on miRNAs as diagnostic biomarkers in PCNSL has not been explored to our knowledge.
Overall, our results indicated miRNAs are suitable as diagnostic biomarkers for PCNSL with high accuracy, with 0.91 for sensitivity, 0.88 for specificity, 70 for DOR, and 0.90 for AUC. The result of the subgroup analysis on the type of specimen indicated the performance of miRNAs in CSF (sensitivity of 0.92 and DOR of 65) was better than that in blood (sensitivity of 0.86 and DOR of 50) for PCNSL detection. However, the combined specificity and AUC of miRNAs in blood for the diagnosis of PCNSL were higher than CSF-based miRNAs assays, where the specificity increased from 0.85 to 0.89, and the AUC increased from 0.88 to 0.94. Thus, both CSF-based and blood-based assays could be considered reliable for clinical application with relatively high diagnostic accuracy. A review  described miRNAs in blood and CSF had important diagnostic advantage, because they were contained in protective vesicles derived from cell membrane and resistant to RNase digestion, exhibiting a remarkable stability. In line with our conclusion, the review showed blood and CSF as relatively noninvasive specimens have been widely used in PCNSL detection.
We conducted another subgroup analysis based on miRNAs profiled (miR-21 and non-miR-21). The performance of miR-21assays for PCNSL detection showed an aggregate sensitivity of 0.89, specificity of 0.89, DOR of 66, and AUC value of 0.95, which were better than non-miR-21 assays (sensitivity of 0.91, specificity of 0.86, DOR of 62, and AUC of 0.91). In agreement with our results, many articles have confirmed the high diagnostic value of miR-21 in both blood and CSF for PCNSL. For example, Baraniskin et al.  were the first to find CSF miR-21 a highly accurate diagnostic marker for PCNSL (AUC 0.94). Another study  indicated a high diagnostic value of serum miR-21 for PCNSL (AUC 0.93). Significant positive correlation of miR-21 was found between serum and CSF in this study (Pearson correlation: r2 = -0.396, p = 0.001). A recent study  demonstrated that miR-21 combined with small nuclear RNA fragments of RNU2-1f in CSF had high diagnostic accuracy, resulting in AUC of 0.987 with a sensitivity of 91.7% and a specificity of 95.7%.
Our study can provide reference to the clinicians regarding the reliability of miRNAs assays at diagnosing PCNSL. Likelihood ratios and post-test probabilities provide information about the likelihood that a patient with a positive or negative test actually has PCNSL or not. In our study, both overall likelihood ratio and post-test probability were moderate (Fig. S1). A positive likelihood ratio of 7 indicates that a person with disease is seven-times more likely to have a positive test result than a healthy person is. Given a pretest probability of 20%, the post-test probability for a positive test result is 65%. Likewise a negative likelihood ratio of 0.11 reduces the post-test probability to 3% for a negative test result.
Two included records [9, 17] showed the diagnostic accuracy of combined miRNAs panels. Baraniskin et al.  reported that combined miR-21, miR-19b, miR-92a analyses yielded a higher discriminatory diagnostic value than any single one. However, the diagnostic accuracy of the combined three miRNAs Zajdel et al.  presented was lower than Baraniskin et al. reported. The reason of this discrepancy most probably lies in the reference groups. The control series Zajdel et al. used comprised patients with benign brain neoplasms and diverse neurological disorders, while Baraniskin et al. series were dominated by multiple sclerosis cases. Further evaluation of the combined three miRNAs with the same reference groups is required to confirm these findings. Zajdel et al.  also revealed that combined miR-let-7b and miR-155 analyses resulted in increased diagnostic accuracy with 96.0% sensitivity and 91.8% specificity. In a word, combined miRNAs panels may improve the sensitivity and specificity of the diagnosis compared with individual one, because one miRNA can target multiple genes and one gene can be regulated by different miRNAs .
According to the QUADAS-2 criteria, all the included records displayed moderate to relatively high quality. In addition, we found no evidence of publication bias. These results strengthened the reliability of our findings.
There are some limitations in this study. First, in view of the rarity of PCNSL, the sample size included was small, which may influence the strength of our analysis to some extent. Further validation based on a larger sample of patients and controls is required. Second, the inconsistent identification of an equal cut-off level for miRNAs in all the studies may have an influence on the final results. Third, we did not further conduct meta-regression analysis to explore whether differences in sample size, miRNAs profiled, and specimen type were the potential sources of the interstudy heterogeneity, because the number of included studies was insufficient.