The effect of metformin on survival of patients with pancreatic cancer: a meta-analysis

We conducted a meta-analysis to analyse the effect of metformin on survival of pancreatic cancer patients at various stages. We performed a systematic search of PubMed, Embase, Cochrane, and Web of Science to identify all relevant studies. Summary hazard ratios (HR) of survival and 95% confidence intervals (95% CI) were calculated with a fixed or random effects model according to inter-study heterogeneity. Nine retrospective cohort studies and two randomized controlled trials (RCTs) were eligible. There was a significant improvement in survival (HR = 0.86, 95% CI 0.76–0.97; P < 0.05) in the metformin group compared with control. Subgroup analysis indicated that metformin improved survival in patients with resection (HR = 0.79, 95% CI 0.69–0.91; P < 0.05) and patients with locally advanced tumors (HR = 0.68, 95% CI 0.55–0.84; P < 0.05) but not in patients with metastatic tumors, even when RCT data were included (HR = 0.99, 95% CI 0.70–1.40; P > 0.05), or were excluded (HR = 0.89, 95% CI 0.61–1.31; P > 0.05). This meta-analysis indicated that the effect of metformin does correlate with tumor stage but should be prudently considered given the limited and variable studies performed to data.


Results
Literature search. The selection process for the literature search is summarized in Fig. 1. Initially, a total of 1134 articles were identified form the search of the four databases. After reviewing titles and abstracts for relevance, 32 articles were evaluated further for eligibility. Of the 32 articles, 21 were excluded: nine articles were conference abstracts without detailed data; five articles pertained to the same study; three articles did not report the outcomes of HR, and only provided the mean survival time; and in another four articles, there were no control groups. Thus, nine cohort studies and two RCTs were included in this systematic review. The blinding method, randomization, generation of random sequences, and allocation concealment of the two RCTs are shown in Supplementary Table S1 and Figure S1.
Quality of studies. The Jadad score 34 of the two RCTs was 5 and 2, respectively. The trial of Reni et al. was of low quality (≤2), and the trial of Kordes et al. was of high quality (>2), according to the Jadad score. The risks of bias in nine retrospective cohort studies are shown in Supplementary Table S2. Of the nine retrospective cohort studies, the NOS scores of five studies were seven, and the remaining four studies were eight. All of the retrospective cohort studies were of high quality (≥7). Study characteristics. The characteristics of 11 included studies are shown in Table 1. In the two RCTs included in this meta-analysis, a total of 181 patients with locally advanced or metastatic pancreatic cancer that could not be treated with resection, were studied. Patients in the study group of Reni et al. were treated with cisplatin, epirubicin, capecitabine, gemcitabine (PEXG) and metformin. In the study of Kordes et al., patients in the study group were treated with gemcitabine, erlotinib, and metformin given orally as a step-up dose. Meanwhile, patients in the control group received gemcitabine, erlotinib, and placebo.
In the nine retrospective cohort studies, one trial did not offer clear definition of metformin exposure, and the other eight trials gave different definitions for metformin exposure. Two trials defined metformin exposure as ever having used it, while one trial defined exposure as ongoing metformin use. In addition, one trial defined metformin exposure as use in the six months before diagnosis, while another trail defined exposure as ever having used the drug in the peri-diagnosis period. Moreover, one trial did not define metformin exposure clearly, yet the entire cohort of patients in the study group was recorded as having been treated with the drug following surgery. The study of Kozak et al. defined metformin exposure as continuous use from the first consultation to discharge after surgery, and the study by Lee et al. defined exposure as cumulative use at one month or longer after diagnosis.
The studies were conducted in four countries: the USA, Italy, Korea, and the Netherlands. In addition, most of these were performed in the last two years. In total, 8089 participants were included in these studies. Among the 8089 participants, 2792 were included in the metformin use group, while 5267 cases were included in the Meta-analysis. The patient survival data are shown in Fig. 2A. All nine cohort studies and two RCTs, including 2792 patients in the metformin use group, reported patient survival as an outcome. The metformin group had significantly better survival compared with the control group (HR = 0.86, 95% CI: 0.76-0.97, P = 0.01). In addition, there was significant heterogeneity between the studies (χ 2 = 23.61, df = 10; P < 0.01; I 2 = 57.6%). Sensitivity analysis was conducted by excluding one study at a time. The study by Lee et al. was identified as largely contributing to heterogeneity. After excluding that particular study, the inter-study heterogeneity weakened (χ 2 = 17.23, df = 9; P < 0.05; I 2 = 47.8%), and the overall result was not affected (pooled HR = 0.89, 95% CI 0.83-0.94; P < 0.01, Fig. 2B).

Subgroup analysis by tumor stage.
Analysis of subgroups by tumor stage was performed. In the subgroup of patients at a resectable stage that included six retrospective cohort studies (4012 patients), a significantly better survival was observed in the metformin group compared to the non-metformin group (HR = 0.79, 95% CI: 0.69-0.91, P < 0.01, Fig. 3A). In the subgroup of patients with local advanced tumors, which included three retrospective cohort studies (548 patients), a significantly better survival was also observed in the metformin group compared to the control group (HR = 0.68, 95% CI: 0.55-0.84, P < 0.01, Fig. 3B). In the subgroup of metastasis patients, which included three retrospective cohort studies and one RCT (319 patients), the difference between the metformin and control group was not significant when the RCT was included (HR = 0.99, 95% CI: 0.70-1. Publication bias. No significant publication bias was found by funnel plots and Begg's test (P = 0.64, Supplementary Figure S2).

Discussion
In this meta-analysis of cohort studies and RCTs investigating metformin in patients with pancreatic cancer, we demonstrated that metformin significantly benefited patients with pancreatic cancer with regard to survival. However, the later the tumor stage, the more obscure the effect of metformin with respect to its efficacy.
The antitumor effect of metformin has been suggested both by epidemiological data and laboratory studies, which provided the rationale for the clinical use of metformin in cancer treatment. Zhang et al. reported a meta-analysis of four cohort studies with significant heterogeneity, and the source of heterogeneity was not further analyzed. In recent years, a number of cohort studies and two RCTs have been conducted on the same subject, and the RCTs showed a different trend with respect to the effect of metformin on pancreatic cancer patients compared with the cohort studies. Therefore, we conducted this new meta-analysis on all nine cohort studies and two RCTs with particular focus on the source of the heterogeneity. The results indicate that tumor stage may contribute greatly to the heterogeneity of the results.
Although the results of the two RCTs were inconsistent with a meta-analysis of previous cohort studies i.e., the benefit of metformin on pancreatic cancer demonstrated in the meta-analysis was not corroborated in the RCTs, this can be explained by the heterogeneity for assigned patients. As shown in Figs 3 and 4, the meta-analysis showed that metformin only benefits patients without metastatic disease depending on whether the RCTs were included, but the assigned patients in both of the RCTs included metastatic pancreatic cancer. cancer patients contributed greatly to the results. Due to the significant heterogeneity between subgroups and the lack of RCTs with resectable pancreatic cancer patients included, we also retrieved clinical trials focusing on the effect of metformin on other solid tumors from PubMed to serve as a reference. Four trials showed that metformin benefited patients with resectable cancer in colorectal cancer 35 , gastric cancer 36 , endometrial adenocarcinoma 37 , and bladder cancer 38 , while one RCT showed no beneficial evidence for metformin on metastatic cancer in non-small cell lung cancer 39 . These findings also suggest that tumor stage plays an important role in the heterogeneity between studies. The exact reason for the heterogeneity between different stages of tumors is unknown. We tentatively speculate that the concentration of metformin in the tumor tissue may play an important role. In patients with metastatic disease, tumor per se was always dense irrespective of large tumor burden, thus the concentration of metformin in neoplastic tissue meant that it was difficult to accumulate drug at a therapeutic level. Therefore, the treatment effect on patients in this group is difficult to explain. However, in patients with resection, plasma metformin might act directly on circulating cancer cells and micro-lesions, which are not macroscopic. Hence, patients in this group had an improved survival with metformin. Furthermore, Kordes et al. and Cerullo et al. also observed an improved survival in patients treated with metformin in a dose-dependent manner, which means that survival benefit increased with a rising daily dose. The improved survival of patients with a higher daily dose may indirectly reflect the influence of insufficient concentration of metformin in neoplastic tissue and its subsequent antitumor effect.
The two RCTs, which are more robust than cohort studies in terms of methodology, showed no beneficial effect of metformin. Due to these disappointing results, which were inconsistent with cohort studies and laboratory studies, the impact of metformin on the survival of pancreatic cancer patients seems very limited. However, the tumor stage varied in clinical trials. Moreover, results of cohort studies were consistent with RCTs, if only metastasis cancer patients were involved. Therefore, the results of the two RCTs are not sufficient to infer that metformin has no beneficial effect on the treatment of patients with pancreatic cancer, especially resectable cancer. Furthermore, perhaps what is more important in the trial by Reni and colleagues is that they evaluated not only the effect of metformin on survival but also the drug's toxicity to patients. Their study demonstrated that metformin combined with chemotherapeutic agents was well tolerated, which has paved the way for subsequent studies.   The use of metformin in pancreatic cancer is similar to that of immunomodulatory treatment of critical illnesses. Both treatments have been documented to be beneficial in animal experiments and primary clinical studies but subsequent randomized clinical trials have demonstrated no beneficial effect or adverse effects, which made the treatments questionable 40 . However, subsequent studies using immunomodulators have indicated that certain populations of patients do benefit from these treatments 41,42 . Similarly, since the trials were heterogeneous with regards to the tumor stage of pancreatic cancer, it is still possible that patients with resectable or locally advanced pancreatic cancer can benefit from metformin. According to the present meta-analysis and systematic review, metformin has the potential to benefit pancreatic cancer patients, especially those patients with tumor at an early stage. Therefore, further clinical trials should focus on patients with resectable cancer.
In interpreting the results of this meta-analysis, both the limited number of available studies and the methodological quality of the included trials should be noted. Of the 11 studies included, only two studies were RCTs, and the methodological quality of one trial was relatively low. Considerable heterogeneity was observed among these trials with regards to the effect of metformin. However, the limited number of studies hampered the power of the subgroup analysis conducted due to heterogeneity. Therefore, there is clearly a need for further investigations that include robust methods and analysis.

Materials and Methods
Systematic literature search. A systematic electronic search was conducted in the PubMed, Embase, Cochrane, and Web of Science databases. The searches were restricted to human disease published before July 23, 2016. The keywords "metformin" and "pancreatic cancer" were used but language restriction was not imposed. Search results were imported into a database and the duplicate records were removed. The titles and abstracts were then screened for relevance. If more than one article was published by the same research team on the same topic, only the most recent article was selected in our study. Next, the full-text papers were checked for eligibility. If the paper was not published in English, we sent an e-mail to the author to ask for an English edition. If an English edition was provided, the study was included. If not, it was eliminated. There was no need for participant consent in this review, since studies included were published without personal identifiable information.
Inclusion and exclusion criteria. We included all human cohort studies and RCTs that investigated the effects of metformin on survival of pancreatic cancer patients. We excluded: (1) cohorts that had no control group, (2) studies that did not report the essential outcomes (survival with HR and 95% CI), and (3) cohorts that included patients with other types of pancreatic tumor (e.g., pancreatic neuroendocrine tumors) and those that did not report the results for pancreatic cancer (adenocarcinoma) separately. The literature retrieval and screening were conducted respectively by two authors (LXG and LT), and a third author (GSM) was consulted when there was uncertainty or disagreement.

Data extraction.
From the included studies, the following information was extracted: design of study, tumor stage, sample size, treatment strategy or exposure to metformin in study group, intervention in the control group, and survival. If the original data for the meta-analysis were not provided directly in the text, then we sent an e-mail to the corresponding author or extracted data from Kaplan-Meier curves in a previously described way 43, 44 . Quality assessment. The internal validity for RCTs was determined using the Jadad scale and risk bias tool of the Cochrane Collaboration. The Newcastle-Ottawa Scale (NOS), including eight quality criteria, was applied to evaluate the quality of retrospective cohort studies. Statistical analysis. All summarized data were analyzed using Stata, Windows version 12.0. The HR with 95% CI was used as a surrogate for survival effect. Heterogeneity between different trials was assessed by the chi-square test, and the extent of inconsistency was evaluated by the I 2 statistic. A Mantel-Haenszel random-effects model was used for data if heterogeneity was significant; otherwise, the fixed-effects model was applied. Sensitivity analysis was conducted to assess the impact of each study on the pooled HR and heterogeneity by removing one study at a time. Publication bias was investigated by funnel plot analysis using Begg's regression test if sufficient studies were included. Two-tailed P-values < 0.05 were considered statistically significant. Data availability. All data generated or analyzed during this study are included in this published article (and its Supplementary Information files).