Diabetes and risk of cancer incidence: results from a population-based cohort study in northern Italy

The study cohort included the inhabitants of the Reggio Emilia province (Northern Italy) who were aged 20–84 on December 31^st, 2009. According to Reggio Emilia diabetes registry (accessed in July 2013), we classified the entire resident population as with or without diabetes (Fig. 1).

The methods applied to set up the diabetes registry have been previously described [16]. In brief, diabetes registry was created by deterministic linkage of six routinely collected data sources through a definite algorithm able to ascertain cases and to distinguish types of diabetes (Type 1, Type 2 and secondary diabetes, i.e. diseases of the exocrine pancreas and drug-induced diabetes), treatment, and care model (for Type 2 diabetes). We used routinely collected health databases covering hospital discharge, drug dispensation, biochemistry laboratory (for glycated haemoglobin), disease-specific exemption, diabetes outpatient clinics activity, and mortality. Women affected by gestational diabetes or those receiving metformin for polycystic ovary syndrome were excluded from the registry. Cases initially notified to the registry through record linkage are retained in case they were clinically confirmed by a diabetologist or other physician.

Resident population was followed up for 4 years. Vital status and migration information were obtained from civil registry. Follow up began on January 1^st 2010 and continued until cancer diagnosis, death, emigration or end of follow up (i.e. December 31^st 2013), whichever came first.

Outcome of interest was cancer incidence rate, using the first incident cancer of each site during follow up and identifying it according to Reggio Emilia cancer registry. Only the first cancer was taken into account to calculate the all-sites cancer incidence; therefore, the sum of number of cases for each investigated cancer site turned out higher. The Reggio Emilia cancer registry includes all malignant cancer cases diagnosed in the Reggio Emilia province between January 1st 1996 and December 31^st 2013. Cancer site was coded according to International Statistical Classification of Diseases and Related Health Problems, 10th Revision. Non-melanoma skin cancers (C44), chronic myeloproliferative disorders, and myelodysplastic syndromes (D45-D47) were ruled out, according to the International rules of cancer registries.

Our analysis considered sex, age and foreign status as covariates.

Baseline characteristics of study population were introduced as absolute numbers and percentages or as medians and interquartile range (IQR) and they were stratified by sex and diabetes status.

We calculated Incidence Rate Ratios (IRRs) and 95% Confidence Intervals (95% CI) using the multivariate Poisson regression model for all-sites and for specific investigated sites. Subjects without diabetes were considered as reference group. We did not define a threshold of significance, thus implying that no calculation of statistical power was made. Moreover we did not propose any statistical test. Confidence interval and p-values were only introduced to provide a measure of the likelihood that the differences were due to chance. P-values for IRR heterogeneity between sex and diabetes were calculated using log-likelihood ratio distribution. As latent cancer might have contributed to develop diabetes, we duplicated some analyses just for subjects with at least 2 years of diabetes duration at baseline, in order to reduce reverse causality. Furthermore, we estimated the IRR and 95% CI according to type of diabetes, diabetes duration at baseline (0–1 years; 2–5; 6–10; > 10), and treatment (the latter only for Type 2 diabetes). All models were adjusted by age at baseline, foreign status and sex (when not stratified). Finally, we compared risk among different types of treatment for Type 2 diabetes, using untreated patients (i.e. with diet-only and physical activity-only programs) as reference, and adding years since diagnosis at baseline (as continuous variable) to other covariates. Analyses were performed by use of STATA statistical package Version 12.0.

As concerns Type 2 diabetes, we had an opportunity of classifying diabetic population according to antidiabetic treatment during 2009. We observed an increased risk related to increased therapy complexity, an indicator of disease severity. Our results were confirmed by the analysis performed among Type 2 diabetes subjects where we use also time since diagnosis as covariate. Unfortunately, we could not define treatment duration and consequently disentangle the insulin effect on cancer initiation and possible masked worse metabolic conditions (indication bias) or possible close monitoring practice (detection bias), as some authors did [17]. Insulin association was consistent with other studies, which suggested a direct role of exogenous insulin and insulin analogues in carcinogenicity [19‐22]. On the other hand, a direct role of insulin was inconsistent with the absence of any increase in cancer risk in Type 1 diabetes patients, who experienced a much longer use of insulin in their life.

The highest IRR was detected in patients treated with both OHAs and insulin. Such therapeutic regimen is usually followed by patients who cannot reach their glycaemic target with the help of just OHAs, often as a preliminary step before initiating sole insulin therapy [23]. However, it can also be used in patients treated with insulin who gain weight or in patients with low compliance to insulin regimen. These patients may all show unstable and high glycaemic values or a worse metabolic state, so hyperglycaemia could amplify the hyperinsulinemia effect, thus increasing cancer risk.

Only a slight excess risk was observed in untreated Type 2 diabetes subjects, in comparison to drug-treated subjects. Such excess risk cannot be due to insulin or OHAs, rather, it might be confounded or mediated by overweight and obesity, which are well- known risk factors for both diabetes and many type of cancers [24].

Our study detected an increasing risk for diabetes duration up to 10 years from diagnosis and a subsequent decrease to moderate-higher risk. Our cohort study, which included prevalent cases of diabetes and incident cases of cancer, possibly minimized the so-called indication bias, which was detected by other studies [17, 18] in which follow up started at diabetes onset and tests recommended after a diabetes diagnosis might have increased the probability of detecting prevalent asymptomatic cancers. A decreased relative risk in the last group of diabetic subjects could be due to higher cancer incidence in people with diabetes, leading to a decreased susceptible population. A similar phenomenon of decreased excess risk has been recorded observed for mortality [25] and for cancer [26].

Strengths are represented by population-based approach thanks to clinically confirmed diabetes registry data, gender approach, including evaluation of heterogeneity between sexes and identification of the type of diabetes, including secondary diabetes, whose causality direction may be difficult to disentangle.

Classification according to patients’ current drug exposure could lead to misclassification. Moreover, we could not evaluate the potential role of some other non-assessed potential confounders, such as different types of insulin used (e.g. insulin analogues vs. human insulin), different mean daily doses of insulin, other drugs taken (e.g., acetylsalicylic acid, statins), glyco-metabolic control, and insulin resistance being present or not.

Nevertheless, the results related to “diet only” and “OHAs only” groups could be unaffected by insulin, and the former also by OHA response. Moreover, our cross-sectional classification of drug exposure was likely to be unaffected by immortal time bias (as we do not set conditions on the exposure duration) and by time-window bias (as we made no use of time dependent variables for exposure).

We did not collect relevant data about other risk factors, in particular positive family history for cancer, smoking, alcohol consumption, physical inactivity, BMI, workplace exposure to toxic substances, etc. We carried out a population-based study and, unlike observational studies of patients using clinical databases, we had limited clinical information on the general population. As regards smoking attitude, our data were confirmed by absence of lung cancer excess. On the other hand, we could not adjust for BMI, although we are well aware that BMI and diabetes could share some mechanisms of cancerogenesis or they could be two rings in the same causal chain. Therefore, further studies should be conducted to analyse possible mediation effect of glucose metabolism in the relationship between BMI and cancer.

As we mainly examined Caucasian race, it was impossible to extend our results to other ethnicities, although our analyses took into account foreign status.

Finally, we considered just 4-year follow up. Such time duration may seem quite short, if compared to time lag from exposure and cancer onset for most cancer sites. Thus, we were mainly concentrated on the effect of diabetes exposure occurred in the years before we started follow up.