Dipeptidyl peptidase-4 inhibitors and the risk of skin cancer among patients with type 2 diabetes: a UK population-based cohort study

Data source

We used the Clinical Practice Research Datalink (CPRD), a representative, electronic health records database containing detailed information for more than 50 million patients seen at over 2000 general practices in the UK.17 In this database, clinical diagnoses are recorded using Read and SNOMED-CT classification system and drug prescriptions are recorded using UK Prescription Pricing Authority Dictionary.17 Importantly, this database also records lifestyle, clinical, and anthropometric variables. These variables have been validated and the data and practices are audited regularly to ensure high quality.18–21

Study population

We assembled two new-user, active comparator cohorts, each investigating a specific skin cancer outcome, from January 1, 2007 (the year the first DPP-4 inhibitors entered the UK market) through July 31, 2019. These cohorts compared initiators of DPP-4 inhibitors (alogliptin, linagliptin, saxagliptin, sitagliptin, vildagliptin) with initiators of sulfonylureas (glibenclamide, gliclazide, glimepiride, glipizide, and tolbutamide). Cohort entry was defined by the date of either the first prescription of a DPP-4 inhibitor or a sulfonylurea during the study period. Patients below 18 years of age, those having concomitant use of the study drugs at cohort entry, and those with less than 1 year of medical history in the CPRD before cohort entry were excluded. The latter served as a washout period to identify new users. We then excluded patients previously diagnosed with any type of skin cancer ever before cohort entry and those with end-stage renal disease as it constitutes a relative contraindication to sulfonylurea use. We also excluded patients with a history of use of the study drugs (DPP-4 inhibitors and sulfonylureas), as well as glucagon-like peptide-1 receptor agonist users, as these drugs share a similar mechanism of action with DPP-4 inhibitors.22 Finally, we excluded patients with less than 1 year of follow-up for cancer latency purposes (ie, lag period). Patients diagnosed with the skin cancer of interest during this lag period were excluded, resulting in two cohorts, one specific to each outcome of interest.

We used sulfonylureas as the comparator as they are widely used second-line to third-line drugs among patients with type 2 diabetes,2 and have not been linked to skin cancer in both clinical23 and laboratory settings.24 We did not use other comparators such as metformin (typically initiated at early stage of the disease), insulin (used at an advanced stage), thiazolidinediones (as they are infrequently used due to their association with serious adverse events), GLP-1 receptor agonists (which have been potentially linked with skin cancers), or sodium glucose co-transporter-2 inhibitors (which are relatively new molecules whose use would have restricted the cohort to 2013 and later, which would be prohibitive due to small sample size).

Follow-up

All patients were followed starting 1 year after cohort entry (ie, after the lag period) until an incident diagnosis of the skin cancer of interest, 1 year after switching to one of the study drugs, death from any cause, end of registration with the general practice, or the end of the study period (July 31, 2020), whichever occurred first. Melanoma diagnoses have been previously validated in the CPRD, with a positive predictive value of 85% compared with medical review.25 On the other hand, CPRD has a better documentation of non-melanoma skin cancer, including basal cell carcinoma and cutaneous squamous cell carcinoma, than the UK national cancer registry.26 27 Moreover, validation studies of UK primary care databases have shown the positive predictive value to be 93%28 and 83%29 for basal cell carcinoma and cutaneous squamous cell carcinoma, respectively.

Potential confounders

We considered a wide range of potential confounders, all measured before or at cohort entry. These included age (modeled using cubic splines with five interior knots to account for a possible non-linear relation with the exposure), sex, lifestyle-related factors (body mass index, alcohol-related disorders, smoking status), calendar year (as a proxy for temporal trends in prescribing and changes in ultraviolet radiation, categorized as 2007–2010, 2011–2014, 2015–2019) and region (as a proxy for exposure to sunlight). We considered known skin cancer risk factors, including precancerous photodermatoses (serving as markers of sun exposure), and use of photosensitizing and immunosuppressive drugs. We also considered diabetes-related variables such as hemoglobin A1c (HbA1c), duration of diabetes (calculated as the time between cohort entry and the earliest of a diabetes diagnosis, use of an antihyperglycemic drug, or an HbA1c value of ≥ 6.5%), as well as microvascular (nephropathy, neuropathy, retinopathy) and macrovascular (myocardial infarction, stroke, peripheral arteriopathy) complications of diabetes. Furthermore, we adjusted for the use of antihyperglycemic drugs even before cohort entry (including metformin, thiazolidinediones, meglitinides, alpha-glucosidase inhibitors, sodium-glucose cotransporter-2 inhibitors, and insulin), common comorbidities (heart failure, cancer, obstructive sleep apnea, osteoarthritis, chronic obstructive pulmonary disease, depression, dyslipidemia, gastrointestinal reflux disease, cardiac arrhythmia, hypertension, hypothyroidism) and co-medications (antihypertensives, antiarrhythmics, antiplatelet agents, statins, non-steroidal anti-inflammatory drugs, corticosteroids, biologics, proton pump inhibitors), and markers of healthcare-seeking behavior (uptake of cancer screening (fecal occult blood testing or colonoscopy, mammography, prostate-specific antigen testing) and vaccinations (including influenza and pneumococcal vaccinations) in the year before cohort entry).

Statistical analysis

We used propensity score fine stratification to adjust for confounding.30 In each cohort, we estimated the predicted probability of receiving a DPP-4 inhibitor versus a sulfonylurea by using multivariable logistic regression conditional on the covariates listed above. After trimming patients in the non-overlapping regions of the propensity score distributions, we created 50 strata based on the propensity score distribution of the DPP-4 inhibitor users. In each stratum, DPP-4 inhibitor users were assigned a weight of 1, while sulfonylurea users were weighted in proportion to the number exposed in the corresponding stratum. This method estimates the average treatment effect among the treated, that is, the DPP-4 inhibitor group.

We used descriptive statistics to summarize the exposure groups’ characteristics before and after weighting. Covariate balance before and after weighting were assessed using standardized differences, with a difference of less than 0.10 indicative of good balance.31 Weighted incidence rates of melanoma and non-melanoma skin cancer, with 95% CIs based on the Poisson distribution, were calculated for each exposure group. Weighted Kaplan-Meier curves were used to display the cumulative incidence of melanoma and non-melanoma skin cancer for the exposure groups over the follow-up period. Finally, weighted Cox proportional hazards models were fit to estimate HRs with 95% CIs of incident melanoma and non-melanoma skin cancer in the respective cohorts, comparing DPP-4 inhibitors with sulfonylureas. We also calculated the number needed to treat/harm for both outcomes after 5 years of follow-up by applying the Kaplan-Meier method.32

Secondary analyses

We conducted four secondary analyses. In the first two analyses, we assessed whether the association varied with time since treatment initiation and cumulative duration of use of DPP-4 inhibitors. In these analyses, the exposure was defined in a time-varying manner, updated every person-day of follow-up. Specifically, the time since initiation was calculated as the difference between cohort entry date and end of follow-up, while the cumulative duration of use was the sum of prescription durations since cohort entry until the risk set event. In the third analysis, we assessed whether the association varied with individual drugs (sitagliptin, alogliptin, saxagliptin, linagliptin, vildagliptin). Lastly, we examined potential effect measure modification on the multiplicative scale by age (<65 vs ≥ 65 years), sex, and use of immunosuppressive drugs. Effect modification was tested by including interaction terms between exposures and these variables in the models.

Sensitivity analyses

We conducted two sensitivity analyses to assess the robustness of our findings. First, we repeated the analysis by lengthening the exposure lag period to 3 years and 5 years to address uncertainties regarding the appropriate duration of cancer latency. Second, we used stabilized inverse probability of censoring weighting to investigate the potential for informative censoring from (1) drug crossover or switching during follow-up and (2) competing risk of death from any cause.33–35 All analyses were conducted with SAS V.9.4 (SAS institute, Cary, North Carolina, USA).

Melanoma outcome cohort

The melanoma outcome cohort included 96 739 new users of DPP-4 inhibitors and 209 341 new users of sulfonylureas (online supplemental figure 1). Before weighting, DPP-4 inhibitor users were more likely to be obese, have a longer duration of diabetes, have microvascular diseases, and enter at a later cohort entry year, than sulfonylurea users. After weighting, all covariates were well balanced, with the standardized difference varying between 0.01 to 0.03 (online supplemental table 1). Over a median follow-up of 2.6 years (IQR 1.1–5.0 years), a total of 634 melanoma events occurred generating a crude incidence rate of 60.5 per 100 000 person-years (95% CI 56.0 to 65.4).

After weighting, the use of DPP-4 inhibitors was associated with a 23% decrease in the incidence of melanoma compared with the use of sulfonylureas (49.7 vs 65.3 per 100 000 person-years, respectively; HR 0.77, 95% CI 0.61 to 0.96) (table 1). The cumulative incidence curves diverged after almost 2 years of use (online supplemental figure 2), with the number needed to treat being 880 at 5 years. The time since initiation analysis revealed no association early in the follow-up period (0–2 years HR 1.24, 95% CI 0.83 to 1.86), followed by HRs below the null value but with wide CIs (2.1–5 years HR 0.59, 95% CI 0.41 to 0.84; >5 years HR 0.69, 95% CI 0.42 to 1.13) (table 1). Similarly, with cumulative duration of use, the HRs decreased with longer durations of use, with at least 5 years of use associated with a 67% decreased hazard (HR 0.33, 95% CI 0.14 to 0.74) (table 1).

The results were consistent across the individual drugs, with HRs below null value with wide CIs, except sitagliptin (the most common DPP-4 inhibitor in the cohort), where the CIs excluded the null value (table 1). The association was not modified by age, sex, or prior use of immunosuppressants (online supplemental figure 3). Overall, the sensitivity analyses aligned with the primary results, although the inverse probability of censoring weighted analysis generated wide CIs (online supplemental figure 4).

Non-melanoma skin cancer outcome cohort

The non-melanoma skin cancer cohort included 96 411 DPP-4 inhibitor new users and 208 626 sulfonylurea new users (online supplemental figure 5). As noted in the melanoma cohort, there were imbalances in terms of obesity, microvascular diseases, diabetes duration, and cohort entry year. After weighting, the covariates were well balanced between the exposure groups (standardized difference between 0.01 to 0.03) (online supplemental table 2). There were 4702 non-melanoma skin cancer events during a median of 2.6 years (IQR 1.1–5.0 years), generating an incidence rate of 455.7 per 100 000 (95% CI 442.8 to 468.9) person-years.

Overall, the use of DPP-4 inhibitors was not associated with the incidence of non-melanoma skin cancer (448.1 vs 426.1 per 100 000 person-years, respectively; HR 1.06, 95% CI 0.98 to 1.15) (table 2). The cumulative incidence curves overlapped throughout the follow-up period (online supplemental figure 6). The HR remained close to the null value in all categories of time since initiation analysis (0–2 years HR 1.09, 95% CI 0.93 to 1.28; 2.1–5 years HR 1.10, 95% CI 0.98 to 1.24; >5 years HR 1.04, 95% CI 0.88 to 1.22). Similarly, there was no consistent pattern with cumulative duration of use, with all HRs around the null value (0–2 years HR 0.98, 95% CI 0.87 to 1.11; 2.1–5 years HR 1.16, 95% CI 1.02 to 1.31; >5 years HR 0.85, 95% CI 0.68 to 1.08).

The analysis stratified by individual drugs generated effect estimates similar to the primary analysis, with CIs including the null value for all drugs (table 2). While there was a 13% increased hazard in the subgroup of patients above 65 years, the CIs overlapped with the estimate for patients below 65 years (online supplemental figure 7). Effect measure modification by sex or immunosuppressant use was also not detected (online supplemental figure 7). Sensitivity analyses resulted in estimates largely overlapping with the primary analysis results (online supplemental figure 8).