The effect of metformin therapy on incidence and prognosis in prostate cancer: A systematic review and meta-analysis

The relationship between metformin and prostate cancer (PCa) remains controversial. To clarify this association, the PubMed, Embase and Cochrane library databases were systematically searched from their inception dates to May 23, 2018, using the keywords “metformin” and “prostate cancer” to identify the related studies. The results included incidence, overall survival (OS), PCa-specific survival (CSS) and recurrence-free survival (RFS), which were measured as hazard ratios (HR) with a 95% confidence interval (95% CI) using Review Manager 5.3 software. A total of 30 cohort studies, including 1,660,795 patients were included in this study. Our study revealed that metformin treatment improves OS, CSS and RFS in PCa (HR = 0.72, 95% CI: 0.59–0.88, P = 0.001; HR = 0.78, 95% CI: 0.64–0.94, P = 0.009; and HR = 0.60, 95% CI: 0.42–0.87 P = 0.006, respectively) compared with non-metformin treatment. However, metformin usage did not reduce the incidence of PCa (HR = 0.86, 95% CI: 0.55–1.34, P = 0.51). In conclusion, compared with non-metformin treatment, metformin therapy can significantly improve OS, CSS and RFS in PCa patients. No association was noted between metformin therapy and PCa incidence. This study indicates a useful direction for the clinical treatment of PCa.

Metformin therapy and PCa recurrence free survival. Figure 4 indicates that RFS was assessed in 8 studies. The HRs for RFS in PCa patients taking metformin compared with those not taking metformin was 0.60, [95% CI: 0.42∼0.87] P = 0.006. Interstudy heterogeneity was noted (I 2 = 63%, P = 0.009). In the subgroup of basic treatment, metformin therapy improved the RFS of PCa patients who accepted radiotherapy (n = 3 HR = 0.41, [95% CI: 0.29∼0.58], P < 0.00001). The subgroup studies consist of study region, sample sizes, diabetic only, study setting and study design (Table 5).
Metformin therapy and incidence of PCa. Figure 5 indicates that incidence of PCa was assessed in 12 studies. The HR for PCa patients taking metformin compared with those not taking metformin was 0.86 [95% CI: 0.55∼1.34], P = 0.51. Interstudy heterogeneity was noted (I 2 = 98%, P < 0.00001). In our subgroup, 6 studies are classified according to their participants' race, including African American, Hispanic/Latino, non-Hispanic white and Asian. Non-Hispanic whites with metformin therapy exhibit a reduced incidence of PCa (HR = 0.86, [95% CI: 0.76∼0.98], P = 0.02). No associations were found between metformin usage and African Americans, Hispanic/Latinos and Asians. The subgroup studies consist of study region, sample sizes, race, duration of metformin therapy, cumulative dose of metformin, Gleason of PCa, advanced PCa, diabetic only and cumulative duration. All subgroup analyses did not reveal any benefits for reducing the incidence of PCa ( Table 6).

Assessment of heterogeneity.
There was evidence of considerable heterogeneity in OS (I 2 = 89%, P < 0.00001), CSS (I 2 = 67%, P = 0.006), RFS (I 2 = 63%, P = 0.009) and incidence of PCa (I 2 = 98%, P < 0.00001). Subgroup analyses investigating potential sources of heterogeneity demonstrated that study region, study design, study setting, sample size and diabetic only were not significantly associated with the heterogeneity in this meta-analysis. We also conducted a sensitivity analysis in which one study was removed at a time and found that the Zaosky 2017 study was the source of heterogeneity in the meta-analysis for RFS. When the study by Zaosky 2017 was removed, the heterogeneity in RFS decreased (I 2 = 57%, P = 0.03), and the results remained stable (HR = 0.66, 95% CI: 0.45-0.96). In the group of incidence, Tseng's 2014 study was the source of statistical heterogeneity in the meta-analysis. When the study by Tseng 2014 was removed, the heterogeneity in incidence decreased (I 2 = 67%, P = 0.0007), and a meta-analysis of incidence results remained stable (HR = 0.96, 95% CI: 0.84-1.10). When the abovementioned studies were removed, the meta-analysis of RFS and incidence demonstrated statistical robustness. No study markedly affected the heterogeneity in the group of OS and CSS. This sensitivity analysis confirms the robustness of our results.
Publication bias. Egger's and Begg's tests revealed the possibility of publication bias for OS (0.677), CSS (0.816), RFS (0.526) and incidence (0.284). No obvious publication bias was noted in our analysis.

Discussion
In the past few years, the controversial results of metformin in the incidence and prognosis of PCa have been increasing. Increasing experimental research reports that metformin exhibits its own advantages in PCa treatment in vitro. Comstock et al. 12 reported that the cyclin D1 pathway was related to PCa cell cycle progression and androgen-dependent transcription. Metformin inhibits PCa cell proliferation by reducing cyclin D1 activity 13 .

Magliano 2012
Cohort *** ** ** 7 Coinciding with the prognosis of PCa, the association between metformin therapy and PCa incidence is controversial. Bansal et al. 10 demonstrated that diabetes reduced the diagnosis of PCa by 14% compared with those without diabetes. Some studies noted that compared with non-diabetic patients, diabetic patients exhibited reduced levels of testosterone, which decreased the incidence of low-grade PCa 11,35-37 . However, Azoulay et al. 38   summarized data from the UK General Practice Research database and found that metformin intake increases the incidence of PCa. Various studies reported conflicting results with the association of metformin usage and PCa diagnosis [39][40][41][42][43] . To clarify this association, we included 12 studies and found no association between metformin usage and the incidence of PCa. A previous meta-analysis provided similar results between PCa risk and metformin exposure 44 . In contrast, two studies 45,46 reported slight reductions (12% and 9%) in PCa risk and metformin use with substantial heterogeneity. (I 2 = 74.7% and 51%). In their meta-analysis, all the included studies were published earlier than 2014. However, our study identified 12 studies and included 1,431,979 male subjects, a larger population group than previous studies 45,46 . Moreover, more than 92% studies were published in the past five years. Therefore, our results gained stronger statistical power. PCa occurrence and outcomes vary considerable between racial and ethnic groups. Siegel et al. 1 reported that PCa incidence and mortality are generally highest among American Africans, whereas Asians exhibited the lowest PCa rates. However, compared with non-Asian patients with type 2 diabetes, Asian patients with type 2 diabetes exhibit a significantly increased risk of PCa 10,47 . One large population-based study reported that Hispanics undergoing metformin therapy exhibited an evident reduction in PCa incidence, whereas metformin usage is not associated with PCa incidence among African Americans and non-Hispanic whites 48 . Previous meta-analyses on this topic revealed no association between metformin and incidence of PCa in either Western-based or Asian-based    49 . However, Western and Asian populations were only classified based on geography, and this studies were limited by significant heterogeneity (I 2 = 88%). Unlike the previous study, we classified all participants as African American, Latino/Hispanic, Non-Hispanic white and Asian. We found that metformin use is associated with a 14% reduction in the risk of PCa among non-Hispanic whites with the presence of heterogeneity (I 2 = 49%). However, metformin therapy did not decrease the risk of PCa among American Africans (I 2 = 26%). A high degree of heterogeneity was noted among Hispanics/Latinos and Asian (94% and 100%, respectively). This high heterogeneity is consistent with a previous study 50 . We found that this evidence heterogeneity is heavily influenced by the studies of Raval and Tseng, which were large studies with an extreme risk estimate. Given that there were only two studies in each subgroup, these studies also had a high level of precision and a high Newcastle-Ottawa score, we included this study. In other subgroup analyses, duration of metformin therapy, cumulative metformin dose, and study region exhibited no association with the incidence of PCa. Moreover, we found that metformin usage was not associated with the Gleason scores of PCa. There was evidence of considerable heterogeneity in OS (I 2 = 89%, P < 0.00001), CSS (I 2 = 67%, P = 0.006), RFS (I 2 = 63%, P = 0.009) and incidence of PCa (I 2 = 98%, P < 0.00001). A sensitivity analysis found that Zaosky 2017 and Tseng 2014 were the sources of heterogeneity in the meta-analysis for RFS and PCa risk. The study by Zaorsky 2017 failed to report the exact start and stop times of metformin. Information on the timing and amount of metformin use was unclear, which might cause a time bias and lead to heterogeneity. In Tseng 2014, no information was available on lifestyle variables, such as smoking status, alcohol consumption, or diet, that potentially influenced the risk of PCa. Although 4 studies 39,41,43,49 also fail to record the lifestyle in the group of incidence.   Differing from these studies which lack information on lifestyle factors, Tseng only focuses on the local Asian administrative databases. We suppose that the lack of information on lifestyle factors in Asian may be the potential reason for heterogeneity in Tseng et al. Moreover, there are only two studies which focus on the Asian in the group of incidence. We use subgroup analysis and found that these two Asian studies also have a high degree of heterogeneity (I 2 = 100%). The participants in Tseng et al. were from the National Health Insurance reimbursement database, which is a local Asian database in Taiwan.While in Chen et al. 49 , the participants were Asian Canadian and come from British Columbia Cancer Agency in Canada. This difference of such criteria included may also lead to the heterogeneity in Tseng 2014. The Newcastle-Ottawa score revealed that these two studies were not more biased compared with other studies and had large sample sizes. Therefore, it is inappropriate to exclude these two studies from our meta-analysis. There were various strengths of our meta-analysis. First, we comprehensively searched relevant studies using Embase, PubMed and Cochrane without publication date or publication type limits by extracting the maximal number of dates in suitable studies. Second, a total of 30 cohort studies including 1,660,795 individuals were included in our studies, which allowed us to quantitatively assess the relationship between metformin intake and PCa. Third, various subgroup analyses, such as PCa treatment, race, duration of metformin therapy, cumulative metformin dose and PCa Gleason score, could provide precise evidence for metformin use in clinical practice. Fourth, we only included the patients with metformin monotherapy and reduced the anticancer bias of other medications.
There were some limitations of our meta-analysis. First, two studies 51,52 did not report the number of metformin users and non-metformin users, and 1 study 53 did not separate type 1 and type 2 diabetes, which may affect the accuracy of the final result. Second, the accuracy of the summary estimates is influenced by different survival analysis methods. Although a multivariate Cox proportional hazards model was used in most of the studies, 2 studies 54,55 did not report their statistical models. Third, because our meta-analysis exclusively focused on studies written in English, a language bias might exist. Fourth, most of our included studies were retrospective studies, which affected the quality of evidence for our meta-analysis.
In conclusion, our meta-analysis suggested that metformin therapy exhibits advantages in improving the prognosis of PCa, but no association was noted between metformin usage and PCa incidence. Moreover, PCa patients with metformin therapy accepting radical radiotherapy exhibited more dramatic effects on OS, CSS and RFS. These studies demonstrated a useful direction for the clinical treatment of PCa. Further randomized controlled trials are needed to confirm the association of PCa and metformin usage.

Materials and Methods
Study selection. Two authors (He & Hu) performed an electronic search of the PubMed, Embase, and Cochrane databases for relevant English studies (the last search update was May 20, 2018). The search strategies included 'metformin' , 'biguanide' , 'Dimethylbiguanidine' , 'Prostate cancer' and 'Prostate Neoplasms' . All the included studies met the following criteria: 1) Study designs must be prospective or retrospective cohort study. Studies must compare metformin users and non-metformin users. 2) Studies must analyse the PCa incidence, overall survival (OS), PCa-specific mortality (CSS) or recurrence-free survival (RFS). We excluded the following types of studies: reviews, case-control studies, studies of interventions other than metformin, articles assessing outcomes following metformin use in animal models, metformin use in other populations, studies including metastatic PCa patients at diagnosis and in vitro studies. Language selection focused on articles written in English. Hazard Ratio (HR) was used as the measure across studies. Given that the PCa incidence was relatively low, odds ratio (OR) were used as an estimate of HR. The prognostic outcomes estimate HRs/RRs with 95% CIs. RFS was defined as the time from the date of PCa patients accepting prostatectomy, radiation therapy or androgen deprivation therapy to the date of biochemical recurrence. After removing duplicate publications, two authors (He & Hu) independently assessed the primary literature by assessing titles and abstracts and then identified the final relevant studies based on eligibility.  PCa incidence and survival estimates were abstracted from the final studies and pooled using a random-effects model. Standard Cochran's Q test and I 2 statistics were used to identify heterogeneity between the included studies. A value of I 2 statistics >50% and p-value <0.1 indicated significant heterogeneity. When heterogeneity was significant, we explored the potential influential variables between included studies and pooled the results into subgroup analyses. Publication bias was detected with the Begg and Egger's regression intercept test by using STATA 13. (Stata Corp LP, college Station, TX).

Ethics approval.
Ethical approval was not sought as the study was based entirely on previously published data.

Data Availability
The study was based entirely on previously published data.  Table 6. Subgroup analysis of Incidence.