Use of insulin and insulin analogs and risk of cancer - systematic review and meta-analysis of observational studies.

The present study was developed according to the PRISMA guidelines [ 18 ], and supplemented by guidance from the Cochrane Collaboration Handbook [ 19 ]. The protocol was registered on Prospero (registration number CRD42012002428) [ 20 ]. The following PICOS eligibility criteria were applied: Population : diabetes patients. Exposure : diabetes patients using any exogenous human insulin or insulin analogues. Comparison : diabetes patients, with or without use of antidiabetic drugs (i.e. use other types of insulin, non-insulin antidiabetic drugs, not use any insulin, or not use any antidiabetic drugs). Studies that only had persons without diabetes as comparator group were excluded. Outcome : incident cancer at specific sites or cancer at any site as a composite outcome. Studies that only report the risk of cancer-related mortality are not included. Study design: cohort and case-control studies. The studies had to report sufficient data for proper evaluation of the study population, exposure, comparator and outcome to be considered for inclusion in the present review. We performed a systematic literature search in 5 databases: Medline at PubMed, Embase, Scopus, Web of Science and The Cochrane Library. The last search was performed on 27 November 2012. The CARING project group concurrently performed a systematic review on risk of cancer in persons with diabetes compared to persons without diabetes [ 21 ]. Records from that review were assessed for inclusion in the present review. The specific search strategy for each database is presented in Supplementary Material 1 . Search terms for diabetes, insulin and cancer (or similar terms) were applied in all searches, while terms for risk or incidence were added in free text searches. For Scopus and Web of Science, free-text searches were used. For Medline, Embase and Cochrane, we used thesaurus (MESH and Emtree terms). In addition, we performed a free text search in Medline, Embase and Cochrane Library limited to references published during the last year in order to identify references not yet indexed with MESH and Emtree terms. Except for limiting the free text search to publications from the last year, no restrictions were used on publication date, language or publication status. ØK and VH developed the search strategy for each database in collaboration with a research librarian. ØK performed the final search in the databases, compiled a mutual reference list for all searches and removed duplicate references. ØK and JSL independently screened title and abstract of records for eligibility, and records identified by either of the reviewers as eligible for inclusion were retrieved in full text. If a conference abstract was deemed eligible for inclusion, a full text article was searched for in databases and included for full text reading if found. ØK and JSL independently assessed the full text records for inclusion and records that ØK and JSL agreed on were included in the review. Disagreements were resolved by discussion and by conferring with a third reviewer (PV). From each study, information was retrieved on risk of cancer, cancer site, definitions of exposure and comparator group (reference), covariates, study design, source population, data sources, and patient characteristics including diabetes type, age group and geographical location (country). Data was extracted by ØK and validated by JSL and disagreements were resolved by discussion. Risk of bias was assessed by the Newcastle Ottawa Scale (NOS) [ 22 ]. All studies were scored by two reviewers (ØK, JSL) and disagreement resolved by discussion and by conferring a third reviewer (PV). The user-defined items required in the NOS score were defined as follows (Supplementary Material 2 ): age was the most important adjustment factor, the exposed in cohorts should be representative of the average “diabetic population using insulin”, minimum average exposure duration was 5 years, and loss to follow-up less than 10%. A conservative approach was chosen if information to score specific items were not available in the article, i.e. no points were given on an item if information was uncertain or missing. Initially, the types of exposure-comparator contrasts and cancer sites examined in records included in the systematic review were assessed by inspecting the summary tables (Supplementary Material 3 ). The contrasts can be categorized as: 1) insulin use versus no insulin use; 2) insulin use versus use of non-insulin antidiabetic drugs; 3) users of insulin A versus users of insulin B; and 4) users of insulin A versus users of insulin B or no insulin. Studies that examined contrast 1 and 2 were included in the pooled analyses while contrast 4 was omitted because of few populations. For contrast 3, glargine insulin users versus non-glargine insulin users was the most frequently used contrast and was included in pooled analyses. Separate pooled analyses were performed for each combination of cancer site and exposure contrast (three selected) that had more than one study population available. One study could contribute more than one population to an analysis, e.g. if the presented risk estimate in the original study was stratified by gender. For studies that published several risk estimates for the same cancer site and exposure contrast (e.g. for different study designs), the following algorithm was applied for choosing which estimate to include (in order of importance): 1) estimates with prior cancer excluded was preferred over estimates adjusted for prior cancer: 2) intention-to-treat analysis preferred over other designs (e.g. as-treated analysis); 3) exposure categorized as exclusive use was preferred (monotherapy, e.g. “glargine only” preferred over “glargine and non-glargine”); 4) estimates without latency period preferred. If no decision could be made from this algorithm, reviewer 1 (ØK) made a final decision on which estimate to include. Estimates from statistical models adjusted for more covariates were preferred. Risk estimates stratified by dose or duration of insulin exposure were not included in pooled analyses. Hazard ratio, incidence risk ratio, rate ratio and odds ratio as summary measures for the risk of incident cancer with 95% confidence intervals were retrieved from each study. These measures were weighted based on the inverse of the standard error of the risk estimator from the individual studies. Chi square test were used to measure heterogeneity across studies. DerSimonian and Laird random effects models [ 23 ] was used in the main analyses regardless of the result of the test for heterogeneity. Additional pooled analyses with a fixed effect model were performed if studies did not exhibit statistically significant heterogeneity. Data were prepared in Microsoft Excel 2010 and analyzed in Stata version 8. Risk of publication bias across studies was assessed by Egger’s regression analysis [ 24 ] in Stata version 8.

In the summary tables all cancer sites are presented together (Supplementary Material 3 ). Several studies have more than one risk estimate presented for each cancer site and exposure contrast because the study reported results for several study designs (e.g. with or without latency period, intention-to-treat and as-treated analyses), or reported both an overall risk estimate as well as risk by strata of dose/duration of insulin exposure. Results of individual studies are presented in Supplementary Material 4 separately for the site-specific cancers examined and for any cancer as a composite outcome. Only the preferred risk estimate for each combination of cancer site and exposure contrast according to the algorithm given in Methods is presented. Cancer at any site and at the following 13 specific sites was examined in more than one study per exposure contrast and was eligible for inclusion in pooled analyses: breast, prostate, stomach, pancreatic, liver, colorectal, colon, rectal, respiratory, bladder, kidney, melanoma, and non-Hodgkin’s lymphoma (NHL). The results for these cancer sites (Supplementary Material 4 ) reveals substantial heterogeneity of results, as point estimates for risk were spread both above and below unity (RR=1) for most cancer sites and exposure contrasts. More consistent results (point estimates) may be present for the exposure contrast insulin versus no insulin for any cancer (3 of 4 populations had point estimate above unity, and with statistical significance), pancreas (7 of 8 populations above unity, 6 significant), liver (5 of 6 populations above unity, 4 significant), stomach (3 of 3 populations above unity, 3 significant), respiratory (5 of 6 populations above unity, 4 significant), bladder (4 of 5 populations above unity, 1 significant), kidney (4 of 4 populations above unity, 2 significant), and prostate cancer (3 of 3 populations below unity, 2significant). For the exposure contrast glargine versus non-glargine insulin use, 6 of 6 populations had risk estimate above unity for prostate cancer but none of the individual risk estimates were statistically significant. 14 cancer sites were only examined in one study per exposure contrast and were not included in pooled analyses: leukemia, Hodgkin’s lymphoma (HL), multiple myeloma, brain, head-neck, skin, testis, ovarian, uterus, cervical, thyroid, oesophagus, gastrointestinal, and lymphoma. Results of these studies are presented in Supplementary Material 5 . In total, 34 studies were included in pooled analyses. Table 3 presents the results of pooled analyses by random effects model for the 14 cancer sites and exposure contrasts with sufficient number of studies (populations). Significant increased risk of cancer for the exposure contrast insulin versus no insulin was found for cancer in pancreas, liver, kidney and the respiratory system, and a marginal significance for stomach cancer. A decreased risk was observed for prostate cancer. Non-significant results were observed for any cancer, bladder, colorectal, colon, rectal, non-Hodgkin’s lymphoma, melanoma and breast cancer. For the exposure contrast insulin versus non-insulin antidiabetic drugs, significant increased risk of any cancer, pancreatic and colorectal cancer was observed, while results for prostate and breast cancer were not significant. Glargine use was associated with a significantly decreased risk of colon cancer compared to non-glargine use breast cancer were marginally significant, while any cancer, pancreatic, liver, bladder, colorectal, respiratory and prostate cancer was not statistically significant. Additional fixed effects models were run for studies that did not exhibit significant heterogeneity (p>0.05, Table 3 ). These analyses gave similar results as the random effects model except for an even higher risk for pancreatic cancer. 8 studies only provided risk estimates by dose or duration of exposure [ 33 , 50 , 52 - 55 , 60 , 66 ] while other studies provided dose or duration risk estimates in addition to average risk estimates. However, pooled analyses by dose or duration was assessed as not feasible because these risk estimates were reported for different cancer sites, exposure contrasts and exposure definitions (e.g. mean or cumulative dose, duration since start exposure or cumulative duration. Dose and duration risk estimates were identified for any cancer, breast, pancreatic, prostate, liver, colorectal, ovarian, lung cancer and lymphoma (Supplementary Material 6 ). Egger’s regression test did not reveal any significant (p <0.05) publication bias for any cancer site.

The results from individual studies in the present review revealed substantial variation in reported risk of cancer associated with use of insulin, and varied by type of comparison group for the insulin users. Many studies are too small to make any firm conclusions. The pooled analyses revealed significantly increased or decreased risk of cancer at several sites for insulin users. However, there were few available studies in each pooled analysis, and subgroup analyses of possible determinants of cancer risk like diabetes type was not feasible. It is imperative to consider the data quality and conduct of individual studies when interpreting these results and the choice of study design in individual studies may have an effect on the estimated cancer risk. Extensive review of the quality of methods, design and conduct of studies was not the aim of the present review. A fit-for-purpose system for evaluating the quality of pharmacoepidemiological studies would be useful in any further evaluation of whether the observed associations can be attributed to issues with study design, analysis and low quality of data.

In the present meta-analysis, insulin exposure seems to be associated with an increased risk of cancer in pancreas, liver, kidney, stomach and respiratory system and decreased risk of prostate cancer, when compared to no insulin use. Compared to use of non-insulin antidiabetic drugs, insulin was associated with increased risk of any cancer, pancreatic and colorectal cancer. For users of glargine insulin compared to users of non-glargine insulin, a decreased risk of colon cancer as well as a marginally significant increased risk of breast cancer was observed. However, the results from individual studies reveal substantial variation in the reported cancer risk for most cancer sites. For 11 cancer sites results were only available in one population per exposure contrast. The importance of assessing dose and duration of insulin use in addition to the average risk has been revealed in several studies observing an increased risk of cancer at different sites even in the initial period after treatment initiation or switch in therapy [ 27 , 40 , 50 ], and the exposure duration may be too short to be a causal factor for the occurrence of cancer. In particular, a substantial increased risk of pancreas cancer is observed and reverse casualty is important to consider for this cancer site. Analyses by duration of insulin exposure reveal specially high risk with shorter durations compared to longer durations [ 27 , 63 , 64 , 68 ]. A similar increased risk is observed in the early period after diagnosis of diabetes [ 63 , 81 ]. This could be a result of diabetes as an early sign of pancreatic cancer (protopathic bias) or ascertainment bias after diabetes diagnosis. Confounding by severity or indication is a concern in pharmacoepidemiological studies, and could be more pronounced when comparing a third-line therapy like insulin to first line therapies like metformin in patients with type 2 diabetes [ 14 ]. Characteristics of populations receiving these two therapies can be substantially different concerning diabetes duration, obesity and other factors. This effect may be less pronounced for use of specific insulin types compared to users of other insulin types, although physician preference for specific insulin types cannot be excluded. Furthermore, a protective effect from metformin use has been reported [ 6 ] and this is important to consider when insulin is compared to metformin or other oral antidiabetic drugs. A few studies presented several results for the same comparison but from different study designs, e.g. intention-to-treat and as-treated analysis, with or without latency period, new user design or “prevalent users design”. This enable assessment of the impact the choice of study design has on results. As an example, Colhoun et al. [ 29 ] reported results for use of “glargine only” and breast cancer risk that were substantially different by study design (range 1.47 to 3.65). Thus, if a different algorithm for selection of estimate to include in the present meta-analysis had been applied, the marginally significant results for glargine use and breast cancer could have been different. This is likely to apply for other comparisons as well. During screening, only 2 randomized controlled trials (RCT) that assessed the risk of cancer in diabetes patients allocated to receive insulins were identified. The Origin trial [ 17 ] included 12,537 people with impaired glucose tolerance or diabetes type 2 for an average follow-up time of 6.2 years to study cardiovascular events as primary outcome. Participants were randomly allocated to receive insulin glargine or standard care and risk of new or recurrent cancer was a secondary outcome. There was no difference in risk of any cancer for the glargine group compared to the standard care group (Hazard Ratio 1.00 [95% CI, 0.88-1.13]). No significant difference in risk was reported for specific cancer sites: breast (1.01 [0.60-1.71]), lung (1.21 [0.87-1.67]), colon (1.09 [0.79-1.51]), prostate (0.94 [0.70-1.26]), melanoma [0.88 [0.44-1.75]) or cancer at other sites [0.95 [0.80-1.14]). A long-term safety study designed to assess ocular complications followed 1,017 persons with type 2 diabetes (82). Participants were randomly assigned to insulin glargine or NPH insulin with a mean cumulative exposure of 4 years. As an additional outcome, malignant neoplasms reported as serious adverse events were assessed and occurred in 51 patients and with relative risk 0.63 [0.36-1.09] for glargine. Risk of benign and malignant neoplasms was 0.90 0.64-1.26. Two meta-analyses of RCT data from manufacturer’s pharmacovigilance databases were also identified. Home et al. [ 15 ] analysed data from 12 phase 2-4 RCTs conducted by Sanofi-Aventis on insulin glargine versus any active comparator (insulin or oral antidiabetics) in type 1 and type 2 diabetes patients. Included studies were between 4 and 52 weeks duration except for the study by Rosenstock et al. [ 82 ] mentioned above, and data in the meta-analysis were primarily driven by those data. 10,880 patients were included and incident malignant cancer occurred in 91 patients with relative risk 0.90 [0.60-1.36] for glargine. Dejgaard et al. [ 16 ] performed a meta-analysis of 21 Novo Nordisk-sponsored RCTs of insulin detemir compared to NPH insulin (16 trials) or insulin glargine (5 trials) in patients with type 1 or type 2 diabetes. RCTs of at least 12 weeks duration were included, with median exposure to insulin of 24 weeks (max 115 weeks) in trials of detemir versus NPH insulin, and 51 weeks (max 64 weeks) in trials of detemir versus glargine. Malignant cancer occurred in 21 of 6,644 patients with Odds Ratio 2.44 [1.01-5.89] for NPH insulin versus detemir, and 16 events in 2,049 patients with Odds Ratio 1.47 [0.55-3.94] for glargine versus detemir. Potential flaws in observational studies of insulin use and risk of cancer have been extensively debated, and the quality of studies included in the present systematic review is a concern. As a measure of the quality of each study, we used the NOS score and most studies could be considered as fair to high quality. However, it can be argued that NOS score is a crude quality measure. Generally, NOS takes into account the quality of the underlying data sources but does not fully account for important issues in pharmacoepidemiological studies, such as definition of drug exposure and time-related biases. For instance, the study by Yang et al. [ 51 ] reported a substantial decreased risk of cancer for insulin users compared to nonusers (HR 0.17 [0.09-0.32]). Potentially serious flaws in the study design have been pointed out [ 13 , 83 ] but the study was nevertheless scored as NOS 9. Potential time-related and other biases of other studies included in the present systematic review have been discussed [ 10 , 12 , 14 ] and these studies also received high NOS scores [ 30 , 33 , 52 ]. Thus, the NOS do not seem to fully reflect important aspects of quality of the studies of the present review and has low granularity to distinguish studies of higher and lower quality. The availability of covariates to adjust for confounding varied substantially in included studies (Table 1 and 2 ). The NOS score does to some extent take into account confounder adjustment, however, adjustment for age and one other factor gave full score on this NOS item. The most important cofounders to adjust for may vary by cancer site and a more thorough assessment of confounder adjustment is desirable. Included studies examined a wide variety of exposures and comparators and this is useful for assessing consistency of the association of insulin and cancer. However, there were too few studies (populations) for most combinations of cancer site and exposure contrast to perform pooled analyses, and additional subgroup or meta-regression analyses could not be performed to assess possible determinants of cancer risk such as diabetes type, gender, age, incident or prevalent insulin use and study design. Egger’s regression test did not reveal any significant publication bias for any cancer site. However, the number of studies in each analysis was low and the test may not have sufficient power to distinguish chance from real asymmetry [ 19 ]. Selective reporting was observed within some published studies as only the analyses with significant results were reported [ 44 , 64 , 79 ].

Associations between diabetes mellitus and increased risk of cancer at several sites have been established [ 1 - 3 ]. It remains unclear whether this relationship between diabetes and cancer is direct, e.g. because of hyperglycemia, or if it is mediated through underlying biologic factors like insulin resistance and hyperinsulinemia, or if it is indirectly linked through common risk factors such as obesity. Insulin is a growth factor, and it is biologically plausible that high levels of endogenous insulin or exposure to exogenous, administered insulin could stimulate neoplastic growth [ 4 , 5 ]. In recent years, several studies have reported modification of cancer risk by use of specific antidiabetic drugs. A decreased risk associated with use of metformin has been reported in meta-analyses while results for thiazolidinedione are not conclusive [ 6 - 8 ]. Results from observational studies published in 2009 raised concerns of a link between insulin use and risk of cancer, but the results of these initial studies were inconclusive and conflicting [ 9 - 11 ]. Publication of many studies assessing risk of cancer at different sites from other data sources has ensued. Several of these observational studies have been hampered by methodological issues and did not take into account dose, duration and timing of insulin exposure or lacked information on important confounders [ 10 , 12 - 14 ]. In addition, most studies have been too small for robust quantification of cancer risk, specially for examining cancer sites individually. The ability to study cancer at specific sites individually is important because cancer is not a homogenous disease and different pathways are involved in the aetiology for different subtypes of cancer [ 2 ]. Existing evidence from randomized controlled trials (RCT) is also limited. Two meta-analyses of RCT data published in the wake of the initial observational studies published in 2009 did not find an increased risk for insulin glargine and detemir [ 15 , 16 ]. However, these studies were rather small for studying a rare event such as cancer and were of limited duration. A larger RCT study with 6 years duration that assessed insulin glargine exposure and had cancer incidence as a secondary outcome reported no increased risk of cancer overall and no significant results for site-specific cancers [ 17 ]. However, the general limitations of RCTs regarding representativeness of the study population apply [ 5 ], and this trial may have been too small to properly quantify risk of cancer at specific sites. Clinical evidence suggests that there may be a link between use of exogenous insulin and risk of cancer at some sites but results are conflicting and inconclusive. The CAncer Risk and INsulin analogs (CARING) project aims to assess possible carcinogenic effects of insulin use combining data from health care databases in six European countries. As part of the CARING project, the present review and meta-analysis was undertaken to summarize published results on the topic. To perform a systematic review and meta-analysis of published cohort and case-control studies that examined the risk of any type of cancer associated with use of exogenous human insulin or insulin analogs in patients with type 1 or type 2 diabetes.