Effect of Diabetes Mellitus on Survival in Patients with Pancreatic Cancer: A Systematic Review and Meta-analysis

Concurrent diabetes has been linked with an increased risk of death in many cancers, but findings in pancreatic cancer have been inconsistent. We performed a systematic review and meta-analysis to assess the effect of diabetes on survival in patients with pancreatic cancer. Of 4, 463 original articles, 41 were included in the review; 29 studies with 33 risk estimates were included in the meta-analysis. In the overall comparison of patients with pancreatic cancer and diabetes with their nondiabetic counterparts, the former had significantly higher all-cause mortality (pooled HR: 1.13; 95% CI: 1.04–1.22). Subgroup analyses showed that diabetes was associated with poor survival in patients with resectable disease (HR: 1.37; 95% CI: 1.15–1.63) but not in those with unresectable disease (HR: 1.07; 95% CI: 0.89–1.29). The HR (95% CI) was 1.52 (1.20–1.93) for patients with new-onset diabetes (≤2 years of diabetes duration) and 1.22 (0.83–1.80) for those with longstanding diabetes (>2 years). Diabetes was associated with higher mortality overall in patients with pancreatic cancer. The effect of diabetes on overall survival was associated with the stages of tumor and the duration of diabetes.

Study selection. Our overall search targeted articles describing studies that met the following three criteria: 1) evaluated any prognostic outcome by DM or glycemic status; 2) evaluated a PC patient population; and 3) contained original data analysis. We included studies evaluating type 1 and/or type 2 DM. To avoid overlapping patient populations, we compared data on recruitment years, data source, and geographic location. Publications with duplicate datasets were triaged by keeping the most recent one, the one with the larger study population, or the one with multivariate-adjusted estimates. Articles that met the above three criteria and reported all-cause mortality or OS were included in our systematic review (Fig. 1). To be included in our meta-analysis, articles had to report a risk estimate (e.g., hazard ratio [HR]) relating preexisting DM to subsequent death by using survival analysis regression models, with an estimate of precision, such as standard error (SE) or 95% confidence interval (CI). Articles with missing risk estimates were also included in the meta-analysis if the risk estimates were generated by author contact. Of the 14 authors contacted, 7 responded, 6 provided additional information. We also conducted a systematic review or meta-analyses on long-term, cancer-specific mortality, disease-free survival (DFS), progression-free survival (PFS), postoperative death rate, and choice of cancer treatment in studies with the relevant information including studies that were excluded from the general review and meta-analysis because of overlapping study durations 36-38 . Data extraction and quality assessment. Each article was abstracted by one author and reviewed by the second author for accuracy. Any disagreement was resolved by consensus. If several estimates were reported in the same article, we chose the most fully adjusted estimate (i.e., multivariate regression was selected over univariate regression, which was selected over unadjusted Kaplan-Meier analysis). If an article reported multiple estimates by subgroup only, these estimates were entered separately into our relevant meta-analysis dataset.
We also extracted information on key indicators of study quality with use of Meta-analysis of Observational Studies in Epidemiology (MOOSE) standards 39 for reporting of meta-analyses of observational studies. From each study, we chose the risk estimates that represented the greatest degree of control for potential confounders. Quality was assessed by using elements of the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement 40 .

Data analyses.
The results of the systematic review were summarized qualitatively. The null hypothesis of "no additional mortality risk in cancer patients with preexisting DM" was tested with use of a nonparametric sign test.
For the meta-analysis, P values quoted at less than the specified threshold were assumed to be at the threshold, resulting in a conservative estimate of the significance level. I 2 and Cochran Q estimates were performed in a heterogeneity assessment 41 . A I 2 value of > 50% or a P value of less than 0.1 represented significant heterogeneity. A DerSimonian-Laird random-effects model (D + L) was used to calculate the pooled HR. Otherwise, an inverse variance fixed-effects model (I-V) was selected. The meta-analysis was performed with use of Stata version 12.0 software (Stata Corp, College Station, TX).
To assess the impact of study quality, we conducted sensitivity analyses. We considered studies to be of higher quality and calculated separate pooled HRs if they were population-based (n = 4) or were adapted from full articles (n = 28), with estimates adjusted for confounders (n = 23) and DM evaluated as the primary exposure variable (n = 10). Publication bias was evaluated by using Begg's funnel plot. We performed the Duval and Tweedie nonparametric trim and fill procedure 42 to further assess potential effects of publication bias. This method considers the possibility of hypothetical missing studies, imputes their HRs, and recalculates a pooled estimate. For all tests, a P value (two-sided) of less than 0.05 was considered statistically significant.
Survival analyses reported various outcomes, including cumulative one-year mortality rates, OS, DFS, and PFS. The studies used a variety of analytic techniques including 9 studies using Kaplan-Meier survival analysis only, 32 using Cox proportional hazards regression analysis (only 29 had exact HRs and 95% CIs), which were included in the following meta-analysis (Supplementary Table 3). The time origin for survival analysis was generally the time of cancer diagnosis, except in the case of treatment or surgical cohorts, for which the time of origin was the beginning of treatment or the date of tumor resection. Most studies were clinic-based design and 4 studies were population-based cohorts (Supplementary Table 3).
Systematic review of evidence. The best evidence from each study is summarized here. Of 9 studies using Kaplan-Meier survival curves with the log-rank tests, DM was associated with decreased survival in 3 studies 33,34,51 and 4 studies 19,52-54 with and without statistical significance, respectively. DM was associated with nonsignificantly increased survival in 2 studies 55,56 .
Overall, DM was associated with increased risk of death in 25 estimates, decreased risk of death in 4 studies, and null effect in 16 estimates. The nonparametric sign test rejected the null hypothesis of equal mortality in patients with and without preexisting DM (P < 0.001).
Meta-analysis on DM and all-cause mortality. The 29 studies in the meta-analysis reported both risk (HR) and precision (95% CI). The descriptive data, adjustment or restriction variables, and major findings from each study are described in Table 1. The results of the meta-analysis are shown in Fig. 2. Preexisting DM was associated with a 13% increased risk of death from all causes in PC patients (HR, 1.13; 95% CI, 1.04-1.22). The pooled HR (95% CI) was 1.37 (1.15-1.63) from 13 studies conducted in patients with resectable disease, 1.07 (0.89-1.29) from 8 studies in unresectable disease, and 1.01 (0.93-1.10) from 10 studies in those with all-stage diseases. The Begg's test and Duval and Tweedie trim and fill procedure showed no significant risk of publication bias (Begg's test P = 0.14; Duvall and Tweedie adjusted HR: 1.10; 95% CI: 1.01-1.20; number of imputed studies = 3) ( Supplementary Fig. 1).   DM and cancer-specific mortality. One study 36 provided adjusted HRs of cancer-specific death in patients who had undergone resection. It showed that patients with DM had a significantly higher risk of cancer-specific mortality compared with their non-DM counterparts (HR: 1.37; 95% CI: 1.00-1.89).

Duration of DM and all-cause mortality.
Six studies 20,27,30,33,37,52 evaluated the association between duration of DM and OS of PC patients. In most studies, two years of diabetes duration was used as the cutoff for defining new-onset and longstanding DM. One study 33 demonstrated that both new-onset and longstanding DM were associated with shorter survival by log-rank test. Three studies 20,37,52 did not find a correlation between new-onset DM and OS in PC patients. Two studies 27,30 indicated that OS duration in PC patients with new-onset DM, but not that in patients with longstanding DM, was significantly shorter than was OS in patients without DM. A meta-analysis of the two studies 27,30 with risk estimates revealed new-onset DM as a significant prognostic factor (HR: 1.52; 95% CI: 1.20-1.93). We did not detect any significant heterogeneity (I 2 = 0%, Q = 0.48; P = 0.49). However, the same prognostic value was not found in longstanding DM (HR: 1.22; 95% CI: 0.83-1.80) (Supplementary Fig. 2).

DM and DFS, PFS.
Two studies 28,38 showed that having DM before undergoing tumor resection was independently associated with poor DFS (pooled HR: 1.54; 95% CI: 1.28-1.85) as well as poor OS after adjusting confounders. No significant heterogeneity was detected in the meta-analysis (I 2 = 14.3%, Q = 1.17; P = 0.28) (Supplementary Fig. 3). Moreover, two studies 17,18 found that in patients with advanced PC who were receiving systemic chemotherapy, PFS did not differ between those with and without DM. DM and treatment selection, postoperative mortality. No significant difference was found in the percentage of surgery between patients with and without DM, although DM patients were more likely recommended for resection 27,34 . Patients with DM had a higher likelihood of developing fistulas and acute kidney injury than did those without DM, but overall complication and severity did not differ between them 59 . The relation between DM and a higher postoperative mortality was not conclusive [59][60][61][62] .

Discussion
Our study demonstrated that preexisting DM in PC patients, compared with their non-DM counterparts, was associated with increased risk of all-cause mortality. The risk of all-cause mortality was higher in patients with resected or resectable tumors than in those with nonresectable tumors and was higher in patients with new-onset DM than in those with longstanding DM. These observations could not be explained by confounding factors, publication bias, or undue influence by a single study. To our knowledge, our meta-analysis is the first exclusive study of the association between DM and PC outcome, even though this topic has been investigated by many individual studies. Our results are not in accordance with a previous meta-analysis of long-term all-cause mortality in cancer patients with preexisting DM by Barone et al. 8 , who found that DM was associated with increased risk of mortality in all cancers (HR: 1.41; 95% CI: 1.28-1.55) but not specifically in PC (HR: 1.09; 95% CI: 0.70-1.69). The discrepancy between the two studies could be partially explained by low power (7%) of the previous study since only four studies in the PC subgroup with a population of 1,681 patients, including 477 DM patients, were analyzed. In the current review and larger-scale meta-analyses, we conducted post hoc power calculations and found that our study had 85% power in demonstrating the association between DM and cancer mortality. Furthermore, we observed that the negative effect of DM on survival occurred primarily in patients with resected or resectable pancreatic tumors (HR, 1.37; 95% CI, 1.15-1.63) but not in patients with late-stage disease (HR, 1.07; 95% CI, 0.89-1.29). These observations support the hypothesis that the previous inconsistent findings between individual studies might be partially explained by the different patient populations involved.
There were several limitations in the literature and thus in our systematic review and meta-analysis. First, studies varied in their inclusion criteria, study population, and adjustment for confounding variables, which may have led to both overestimations and underestimations of risks. Nevertheless, our sensitivity analyses, excluding studies that did not adjust for potential confounders, did not materially change the results. Residual or unknown confounding is still possible after adjusting for most relevant confounding factors. The association may not necessarily be causal as well, particularly in the observational studies 63 . Given that the start time of survival of each study differs by cancer stage (ie. tumor resection), we also performed a subgroup analysis by cancer stage (resected/resectable, unresectable and mixed stages), which confirmed the association of diabetes with poor survival in patients with resected/ resectable tumor.
Second, the status of DM ascertainment varied across studies, and the duration of DM was not directly reported in some studies. Moreover, in most studies, diabetic status was based exclusively on past medical history; thus, there was a chance of misclassification, which may have led to underestimation of the number of patients with DM and of the effect of the disease.
Third, overlap in patient enrollment between some studies may have elevated their weight in the quantitative analysis. For example, the patient cohort used by Cannon et al. 28 to derive and test the survival prognostic scoring model included the 209 patients from Chu et al.'s prior study 27 , and these patients were randomly divided into training and validation sets. In addition, two studies conducted at the same institution, i.e., Li et al.'s study of all-stage patients recruited between 1999 and 2008 and Sahin et al.'s study of patients who had undergone resection and were recruited between 1996 and 2001, may have some overlapping patients. However, we expect that the effects of this overlap on the final results of our analysis to be minimal.
The fourth limitation was that most of the articles did not report the types of anticancer and antidiabetic therapies used or their effects on outcomes. This is important because studies have shown that some therapies (e.g., surgery, adjuvant chemotherapy, and the antidiabetic drug metformin) have a more positive effect than others on cancer outcome 64,65 .
There are several potential explanations for the observed association between decreased survival time and DM in PC patients. First, DM may enhance tumor progression via the mechanisms of insulin resistance and inflammation, i.e., the same mechanisms that contributed to the higher risk of PC may also be responsible for the poor outcome of the disease. It has been suggested that hormonal or metabolic abnormalities, such as hyperinsulinemia or hyperglycemia, may affect tumor biology at multiple stages, including malignant transformation, growth, and metastasis 66 .
Second, PC patients with DM may be given less vigorous anticancer regimens because they generally have more contradictions to surgery, chemotherapy, and other treatments, although selection bias associated with types of surgery, chemotherapy, and radiotherapy were not seen in previous studies 27 . The metabolic abnormalities associated with DM may have an adverse effect on response to cancer treatment 67 .
Finally, the high mortality rate observed among patients with DM may partially be due to noncancerous factors, such as complications of long-term DM. However, since most PC patients die of the disease within a short period of time, it is unlikely that the complications of long-term DM would make Scientific RepoRts | 5:17102 | DOI: 10.1038/srep17102 a significant contribution to PC mortality. Only one study measured PC-specific mortality and observed a significant impact of DM 36 .
Previous studies 2,6 have shown that new-onset DM, compared with longstanding DM, was associated with greater risk of PC because new-onset DM was a manifestation of subclinical PC. The main implication of our study is that DM is significantly associated with adverse outcome in PC. Subgroup analyses showed that the effect of diabetes on overall survival was associated with the cancer stages and the duration of diabetes. Our results reveal the need for further prospective studies to confirm DM as a prognostic factor and to assess the possibility of an antidiabetic regimen in the treatment of PC.