This meta-analysis was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (34). A complete network research was done by electronic databases PubMed, Embase, and Web of Science (up to July 20, 2021) through different combinations of the search items: “CD9,” “MRP1,” and “cancer”/”tumor”/”neoplasm”/”carcinoma” and the following limits: human subjects and reports in the English language. All probably appropriate articles were selected, and their references were prudently browsed to recognize else applicable essay. While several researches of the same patient population existed, we involved the available data with the largest sample size.

Reports that meet the following criteria were considered eligible: a) calculated the expression of CD9 for predicting outcome (overall survival (OS) or time to progression (TTP)) in human cancer, b) provided hazard ratios (HRs) with 95% CIs or sufficient information that could estimate it, and c) classified CD9 expression as “high” and “low” or “positive” and “negative.”

The literatures were excluded based upon the subsequent standard: a) researches were published as laboratory paper, case reports, letters, editorials, and abstracts, as well as reviews and proficient opinions; b) experiments were done in vitro or in vivo and not based upon patients; c) literatures without information on HRs, 95% CI of OS and/or disease-free survival (DFS), or even the Kaplan–Meier (K-M) survival curves; d) researches that described the survival outcome of other indicators; and e) unpublished studies.

Two reviewers extracted data from the articles independently and carefully using a standardized form. The third author would independently extract data from the original article. If disagreements exist, consensus would be reached through debate. This article was founded on OS and TTP. DFS and time to recurrence (TTR) that had similar definition were merged into TTP of tumor. The following information was extracted from selected papers: the first author, publication year, country of participants, number of patients considered, kinds of cancer, cutoff values, OS, TTP, Newcastle–Ottawa scale (NOS) score. The chief characteristics of these researches are shown in Table 1 . For some articles, HR can be directly obtained; for the researches in which survival data are presented only with K-M curves, Tierney’s method was used to calculate the HR and 95% CI (46). NOS was employed to evaluate the eligible literature. The scores of eligible articles vary from 6 to 9, which mean that the methodological quality of these papers was high.

Stata14.0 (StataCorp, College Station, TX) was used to perform the statistical analysis. Initially, we calculated the connection between MRP-1 and endpoints (OS and TTP) though pooled HRs and 95% CIs. Cochran’s Q test and Higgins I-squared statistic were applied to evaluate whether heterogeneity was present between selected articles. Heterogeneity was considered significant when p < 0.1 or I² > 50% (47), and then a random-effects model was employed to pool the HRs and 95% CIs; otherwise, a fixed-effects model was used (35). Moreover, subgroup analysis was used to explore the source of heterogeneity. Begg’s and Egger’s tests were employed to discover the potential publication bias. When publication bias indeed exists, the Duval and Tweedie trim and fill method (37) was exploited. For the robust estimation of the results of statistical analysis, sensitivity analysis was performed by removing the original study one by one. The p-value for all tests was two-tailed, and p < 0.05 was defined as statistically significant, except for heterogeneity.

The particular process of searching and filtering is displayed in Figure 1 . We preferentially retrieved 364 records from PubMed, Embase, and Web of Science in accordance with the criteria mentioned previously. Among them, 117 duplicate reports were removed. After the abstracts or full text was scanned, 222 records were excluded by reason of no relevant information provided (n = 197); experimental research (n = 3); without prognosis data (n = 2); and irrelevance (n = 20). Ultimately, 25 records including 5,555 participants were involved in the meta-analysis. The median specimen size was 222, ranging from 34 to 1,349. Among all cohorts, three studies estimated breast carcinoma (21–23), four studies estimated lung cancer (17–20), three studies evaluated colon cancer (24–26), one study estimated clear cell renal cell carcinoma (27), one study estimated malignant pleural mesothelioma (28), one study assessed laryngeal squamous cell carcinomas (29), one study assessed GIST (30), one study evaluated gallbladder cancer (38), one study evaluated gastric GIST (39), two studies investigated oral squamous cell carcinoma (40, 41), one study investigated gastric cancer (42), two studies investigated squamous cell carcinoma of the head and neck (43, 44), one study investigated urothelial carcinoma of the bladder (45), one study investigated endometrial cancers (48), one study investigated esophageal squamous cell carcinoma (49), and one study investigated pancreatic cancer (50). Five studies (25%) were focused on Asians and 20 (75%) on Caucasians. Overall, 17 papers reported on OS and 16 papers on TTP.

In this meta-analysis, we estimated the correlation between CD9 and prognosis of tumor. As is shown in Figure 2 , the result of heterogeneity test illustrated that p < 0.001 and I² = 68.2%; therefore, a random-effects model was applied to evaluate the pooled HR and 95% CI of OS. The outcome implied that declined expression of CD9 was positively correlated with poor OS in solid tumor (pooled HR = 1.88, 95% CI = 1.45–2.43, p = 0.000). As is displayed in Figure 3 , a random-effects model was employed to calculate the pooled HR and 95% CI of TTP as well; the heterogeneity test reported p < 0.001 and I² = 81.7%. The results indicated that low expression of CD9 was significantly associated with shorter TTP (pooled HR = 2, 95% CI = 1.38–2.88, p = 0.000). To explore the source of heterogeneity, subgroup study was therefore executed. We discovered that decreased expression of CD9 was connected to shorter OS in Asians participants (HR = 1.96, 95% CI = 1.49–2.58, p < 0.001; random effects: I² = 66.5%, p < 0.001), as well as in lung cancer (HR = 1.93, 95% CI = 1.18–3.17, p < 0.001; random effects: I² = 79.3%, p = 0.002) and head and neck cancer (HR = 1.98, 95% CI = 1.39–2.82, p < 0.001; fixed effects: I² = 0%, p = 0.89). The correlation was also detected between decreased expression of CD9 and poor TTP in either Asians patients (HR = 1.38, 95% CI = 1.17–1.62, p < 0.001; random effects: I² = 80.4%, p < 0.001) or Caucasian patients (HR = 1.87, 95% CI = 1.41–2.48, p < 0.001; random effects: I² = 85.4%, p < 0.001), and in patients with head and neck cancer (HR = 2.31, 95% CI = 1.59–3.36, p < 0.001; fixed effects: I² = 11.5%, p = 0.323). No other significant connection between CD9 and the two endpoints was detected in the other subgroup analyses.

Begg’s funnel plot and Egger’s test were applied to measure the publication bias of the literature. As is shown in Figure 4A , Begg’s and Egger’s test scores of OS were correspondingly p = 0.869 and p = 0.008. The funnel plot for the OS showed asymmetry. At the same time, the calculation of the TTP yielded publication biases (Begg’s test, p = 0.86 and Egger’s test, p = 0.024) ( Figure 4B ). Consequently, the trim and fill method was employed to make the pooled HR more dependable, and the result showed that the pooled p-value was also less than 0.01 (figure not shown). Furthermore, sensitivity analysis was performed by removing one study in turn, which revealed that no single study would significantly affect the pooled HRs of OS and TTP ( Figure 5 ).