The study-related subject terms and free terms were first identified in PubMed, and then a literature search was conducted in PubMed, Web of science, Science direct, Embase, Cochrane, and Medline databases from the time of construction to July 1, 2023 for the following search terms “Lung Neoplasms”, “Pulmonary Neoplasms”, “Lung Cancer”, “Pulmonary Cancer”, “Cancer of the Lung”, “Cancer of Lung”, “MicroRNAs”, “MicroRNA”, “Primary MicroRNA”, “miRNA”, “Small Temporal RNA”, “RNA, Long Noncoding”, “lncRNA”, “Long Non-Translated RNA”, “Long Noncoding RNA”, “Long ncRNAs”, “Long Untranslated RNA”, “Long Intergenic Non-Protein Coding RNA”, “RNA, Circular”, “circRNAs”, “Closed Circular RNA”, “Circular Intronic RNA”, “ ciRNA”, “Exosomes”, “Diagnosis”, “Diagnoses” and “Diagnoses and Examinations”.

The retrieved literature was summarized and after deleting duplicates and literature for which full text could not be found, irrelevant studies were eliminated by reading the titles and abstracts of the literature, and then screened according to the inclusion and exclusion criteria to finally obtain the target literature.

Inclusion criteria: (1) publicly published diagnostic tests on the accuracy of exosomal miRNA, lncRNA and circRNA in diagnosing lung cancer; (2) the language of the literature was English; (3) the case group was patients with pathologically diagnosed lung cancer; (4) bodily fluids were the criteria for the study samples; (5) the study mentioned the differences in the expression of exosomal miRNA, lncRNA and circRNA in the study controls; and (6) the number of true-positive, false-negative, pseudo-positive, false-positive and true-negative numbers for the diagnosis of lung cancer could be obtained in the study, either directly or indirectly, for exosomal miRNAs, lncRNAs and circRNAs.

Exclusion criteria: (1) duplicate publications; (2) reviews, conference papers, abstracts, lectures, reviews, and case reports; (3) literature that could not be extracted from the four-compartment table; and (4) literature that was co-diagnosed with other tumor markers.

After completing the article retrieval and screening steps, the researcher extracted data from the included literature and performed quality assessment. Data not available directly was harvested using Engauge Digitizer. The QUADAS-2 tool in Review Manager 5.4 software was used to assess the quality of the included literature. QUADAS-2 includes four components: selection of cases, trials to be evaluated, gold standard, case flow, and progress, and articles were evaluated mainly in terms of risk of bias and clinical applicability.

Baseline data indicators include name of included authors, year of publication of the article, country of authors, sample size of included samples, source of cases, source of controls, type of samples, method of exosome extraction, selection of miRNAs, lncRNAs and circRNAs.

Outcome metrics include sensitivity(Sen), specificity(Spe), area under the curve(AUC), number of true positives(TP), false positives(TP), false negatives(FN), and true negatives(TN) of exosomal miRNAs, lncRNAs, and circRNAs used for the diagnosis of lung cancer patients in the included studies.

The miRNA downstream target genes were predicted by miRDB(https://mirdb.org/mirdb/), mirDIP(https://ophid.utoronto.ca/mirDIP/), miRWalk(http://mirwalk.umm.uni-heidelberg.de/) and TargetScanHuman8.0(https://www.targetscan.org/vert_80/).The GSE33532 GeneChip data were downloaded from Gene Expression Omnibus(GEO) database(https://www.ncbi.nlm.nih.gov/geo/), and the predicted genes were screened for the intersection of the predicted genes and differentially expressed genes(DEGs), and the DAVID database(https://david.ncifcrf.gov/home.jsp) was applied for the Gene Ontology(GO) analysis and Kyoto Encyclopedia of Genes and Genomes(KEGG) analysis, Heatmap was plotted by https://www.bioinformatics.com.cn (last accessed on 10 July 2023), an online platform for data analysis and visualization.

Threshold effect analysis was performed on the included literatures, that is, Spearman correlation coefficient test between the logarithm of sensitivity and the logarithm of 1-specificity. If correlation analysis P>0.05 indicates that there is no threshold effect causing heterogeneity. Cochran-Q test and I² test included heterogeneity of non-threshold effects between studies. P ≤ 0.1 and I²≥50% indicated significant heterogeneity; P>0.1 and I²<50% indicated insignificant heterogeneity. The evaluation indexes of combined diagnostic tests included sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, area under the curve and diagnostic odds ratio. Meta-regression and subgroup analysis were used to explore the sources of heterogeneity. The corresponding forest plot of the evaluation index was drawn, the publication bias was assessed by Deek’s funnel plot, and the clinical effect was assessed by Fagan plot and likelihood scatter plot. The software used in this study included Meta-disc 1.4 and Stata 17.0.

The steps for selecting eligible studies are shown in Figure 1 . A comprehensive literature search in six databases identified 422 articles, 294 articles were reviewed after excluding duplicates and those for which the full text could not be found, and 41 articles were finally included after reading the title, abstract, and study content for screening. Therefore, 41 eligible articles were included in the meta-analysis.

Forty-one articles including 68 studies were included. Of these, 28 articles on microRNAs (16–43), comprising 45 studies; 8 articles on lncRNAs (44–51), comprising 9 studies; and 5 articles on circRNAs (52–56), comprising 14 studies. Diagnostic effect data included TP, FP, FN, TN, Sen, Spe, AUC, diagnostic likelihood ratio positive(DLR+), diagnostic likelihood ratio negative(DLR-), and diagnostic OR(DOR). At the same time, in order to ensure that the non-coding RNA studied was derived from exosomes, the characteristics of exosomes were identified by nanoparticle tracking analysis(NTA), transmission electron microscope(TEM) and exosome marker proteins. The information of exosomes in various studies were summarized in Supplementary Table 1 .

Most of the articles provided and specificity information, which could be obtained directly, and the sensitivity and specificity of the other 9 articles were obtained by reading ROC curves with the digitizing software engage Digitizer (18, 23–25, 36, 41, 52–54). The results are shown in Table 1 and Supplementary Table 2 .

The results of the QUADAS-2 quality assessment of the 41 included articles are shown in Supplementary Table 2 and Supplementary Figure 1 , and the QUADAS scores are summarized in Table 1 . Each entry was evaluated with a “yes,” “no,” “unclear” evaluation, with “yes” scoring 1, “no” 1, and “unclear,” 0. We defined studies with a score greater than 4 as high quality studies and those with a score less than 4 as low quality studies. The overall quality of the included literature was good, with 40 high-quality and 1 low-quality literature.

Analysis of the data included in the studies of exosomal miRNAs, circRNAs, and LncRNAs by Meta DiSc14.0 software yielded Spearman’s correlation coefficients between the logarithm of sensitivity and the logarithm of 1-specificity of 0.06, -0.08, and 0.40, respectively, with P=0.689, 0.847, and 0.161>0.05, which implying that there is no threshold effect in the study of exosomal miRNAs, lncRNAs, and circRNAs.

Heterogeneity among the included studies was evaluated using the Cochran Q test and I² value in Stata software. The results of the miRNAs and circRNAs studies were Cochran Q=182.85(P ≤ 0.001), I² = 99%(95%CI,98~99); Cochran Q=195.93(P ≤ 0.001), I² = 99%(95%CI,98~99) suggesting the presence of significant heterogeneity arising from non-threshold effects, while the lncRNAs results Cochran Q=1.38(P=0.251), I² = 0%(95%CI,0~100) indicated better homogeneity.

A random-effects model was used to estimate the diagnostic effects of exosomal miRNAs, lncRNAs, and circRNAs on lung cancer. MiRNA study data had a combined sensitivity and specificity of 0.83(95%CI,0.80~0.86) and 0.83(95%CI,0.79~0.87). Combined DLR+ and DLR- were 4.94(95%CI,3.85~6.32) and 0.20(95%CI,0.17~0.24). The combined diagnostic score and diagnostic advantage ratio were 3.19(95%CI,2.82~3.56) and 24.29(95%CI,16.83~35.04), respectively, and the forest plot is shown in Figure 2 . The included studies showed a significant difference in sensitivity(P ≤ 001, I² = 80.12%),specificity(P ≤ 0.001,I² = 90.79%),DLR+(P ≤ 0.001, I² = 90.07%),DLR-(P ≤ 0.001, I² = 81.91%), significant heterogeneity in diagnostic scores (P ≤ 0.001, I² = 83.90%) and diagnostic superiority ratios (P ≤ 0.001, I² = 100%). And SROC curves were constructed, as shown in Figure 2 , with a combined AUC of 0.90(95%CI,0.87~0.92). The results of these pooled analyses suggest that exosomal miRNAs have good diagnostic efficacy for lung cancer.

The combined sensitivity and specificity of lncRNA study data were 0.72(95%CI,0.67~0.78) and 0.79(95%CI,0.75~0.82). The combined DLR+ and DLR- were 3.42(95%CI,2.77~4.24) and 0.35(95%CI,0.28~0.44). The combined diagnostic score was 2.28(95% CI,1.88~2.68) and the superiority ratio was 9.77(95%CI,6.52~14.63), and the forest plot is shown in Figure 3 . There was significant heterogeneity among the included studies in terms of sensitivity (P ≤ 0.001, I² = 65.22%), DLR-(P=0.01, I² = 63.56%) and dominance ratio (P ≤ 0.001, I² = 98.49%). The SROC curve was constructed, and the combined AUC was 0.82(95% CI,0.79~0.85). These pooled analyses suggest that exosomal lncRNA has a good diagnostic effect on lung cancer.

The combined sensitivity and specificity of the circRNA study data were 0.79(95%CI,0.67~0.87) and 0.81(95%CI,0.74~0.86), respectively, and the forest plot is shown in Figure 4 . The combined DLR+ and DLR- were 4.12(95%CI,3.19~5.32) and 0.26(95%CI,0.17~0.41). The combined diagnostic score and diagnostic advantage ratio were 2.75(95%CI,2.24~3.26) and 15.67(95%CI,9.43~26.06). The included studies showed a significant difference in sensitivity (P ≤ 001, I² = 93.12%), specificity(P ≤ 0.001, I² = 90.29%), DLR+(P ≤ 0.001, I² = 80.09%), DLR-(P ≤ 0.001, I² = 92.26%), significant heterogeneity in diagnostic scores(P ≤ 0.001, I² = 79.63%) and diagnostic superiority ratio (P ≤ 0.001, I² = 100%). And SROC curves were constructed, as shown in Figure 4 , with a combined AUC of 0.86 (95% CI,0.83~0.89). The results indicate the diagnostic value of exosomal circRNAs in lung cancer.

Due to the existence of heterogeneity, meta-regression analysis and subgroup analysis were performed on miRNA and circRNA inclusion in the literature in terms of sample size, sample type, control source, exosome extraction method, and miRNA detection profiles, and the results are shown in Table 2 . The meta-regression results of the study showed that the heterogeneity among the studies of exosomal miRNAs for lung cancer diagnosis was related to the methods of RNA detection(P=0.023), while sample size(P=0.256), types of contrast(P=0.468), sample types(P=0.107), exosome extraction methods(P=0.901), and number of miRNA assays(P=0.232) did not have a significant effect on the combined results. The results of subgroup analysis showed that exosomal miRNAs from plasma were more sensitive than miRNAs from serum extracted exosomes, miRNAs detected by RNA-seq had better diagnostic performance than miRNAs detected by qRT-PCR and the combined efficacy of multiple miRNAs in the diagnosis of lung cancer was good compared to single miRNAs.

LncRNA was included in the literature for subgroup analysis to assess the effect of each variable on the results, and the variables of sample size (P=0.657), sample type (P=0.575), source of control (P=0.396), and exosome extraction method (P=0.350) were not associated with the results of lung cancer diagnosis by lncRNA.

The inter-study heterogeneity of circRNA for lung cancer diagnosis was not related to sample size (P=0.661), sample type (P=0.661) and RNA detection methods(P=0.141), and its heterogeneity may come from methodological heterogeneity and clinical heterogeneity or the selection of fewer variables in this study. Subgroup analysis showed that sample sizes less than 100 had better sensitivity and the diagnostic efficacy of circRNA detected by RNA-seq was better.

The publication bias of the included studies was tested using Deek ‘s funnel plot asymmetry test, as shown in Supplementary Figure 2 . CircRNA-P value was 0.54 and lncRNA-P value was 0.85, which were greater than 0.05, indicating that there was no potential publication bias. miRNA-P value was <0.01, which indicated the presence of publication bias.

To explore the clinical significance of miRNAs, lncRNAs and circRNAs for diagnosing lung cancer, we constructed Fagan plots to interpret the pre-test probability, likelihood ratio and post-test probability. As shown in Supplementary Figure 3 , for exosomal miRNA, the assumed pre-test probability was 20%, and based on the positive diagnostic likelihood ratio, the post-test probability was 55%, indicating that exosomal miRNA has good clinical diagnostic value for lung cancer. For exosomal lncRNA assuming a pre-test probability of 20% and a post-test probability of 46% based on the positive diagnostic likelihood ratio, this indicates that exosomal lncRNA has good clinical diagnostic value for lung cancer. For exosomal circRNA, a pre-test probability of 20% was assumed, and based on the positive diagnostic likelihood ratio, the post-test probability was 51%, indicating that exosomal circRNA has good clinical diagnostic value for lung cancer.

Likelihood ratio scatterplot analysis was further performed. The scatterplot was divided into 4 quadrants, i.e., upper left limit (LUQ), upper right limit (RUQ), lower left limit (LLQ) and lower right limit (RLQ). In the LUQ, DLR+ is greater than 10, DLR- is less than 0.1, indicating that the test can confirm and exclude lung cancer. In the RUQ, DLR+ greater than 10, DLR- greater than 0.1 only confirms lung cancer. In the LLQ, DLR+ less than 10, DLR- less than 0.1 only excludes lung cancer. In RLQ, DLR+ less than 10, DLR - greater than 0.1 suggests that neither lung cancer can be confirmed nor excluded. As shown in Supplementary Figure 4 , the diagnostic ability of exosomal circRNAs was limited in clinical confirmation and exclusion of lung cancer. miRNAs and LncRNAs had a stronger diagnostic ability than circRNAs.

To explore the function of miRNAs with potential clinical value, we predict the downstream target genes of miRNAs. From the miRWalk, miRDB, TargetScan, and mirDIP databases, we predicted 3100 possible miRNA downstream target genes ( Figure 5A ), and 609 differentially expressed genes obtained from downloaded GSE33532 genechips, we found 125 target genes ( Figure 5B ), as shown in Supplementary Table 3 . GO analysis ( Figure 5C ) shows that These miRNA-targeting genes are involved in the positive regulation of cell migration, nervous system development, and positive regulation of transcription from RNA polymerase II promoter, cellular response to hormone stimulus, response to cold, female pregnancy, brain development, regulation of cell effort, transcription from RNA polymerase II promoter and axon guidance process. KEGG analysis ( Figure 5D ) showed that these miRNA-targeted genes were involved in cAMP signaling pathway, proximal tubule bicarbonate reclamation, renin secretion, bile secretion, cell adhesion molecules, TGF-beta signaling pathway, axon guidance, focal adhesion, chemical carcinogenesis-receptor activation, MAPK signaling pathway.