Liraglutide use and evaluation of pancreatic outcomes in a US commercially insured population

Abstract Aims Both acute pancreatitis (AP) and pancreatic cancer (PC) have been areas of focus for studies of incretin drugs. This 5‐year prospective cohort study aimed to quantify possible associations between liraglutide and risk of AP and PC as compared to other antidiabetic drugs (ADs). Materials and methods Patients initiating liraglutide or other ADs who were enrolled in a US health plan (2010‐2014) were included. Comparisons of AP and PC incidence rates were made between matched cohorts of liraglutide initiators and initiators of other ADs. Adjudicated AP cases and algorithm‐based PC cases were identified. Propensity score‐matched intention‐to‐treat (ITT) and time‐on‐drug (TOD) analyses were completed using Poisson regression. A latency analysis was performed for PC. Results Median follow‐up was 405 days for AP cohorts (9995 liraglutide, 1:1 matched to all comparators) and 503 days for PC cohorts (35 163 liraglutide, 1:1 matched to all comparators). In the primary AP analysis, “current” use of liraglutide was not significantly associated with elevated risk across comparators (all comparators relative risk [RR] = 1.2; 95% confidence interval [CI], 0.6‐2.3). ITT results were similar where, in the primary analysis, no RRs were significantly associated with PC (all comparators RR = 0.7; 95% CI, 0.3‐1.4); latency and TOD analyses did not alter findings. There was no evidence of a dose‐response effect. Conclusions Liraglutide was not associated with an increased risk of AP or PC, although risk estimates were more variable for AP, and numbers of cases for both outcomes were limited because of the rarity of outcomes.


| INTRODUCTION
Approximately 30 million people have diabetes in the United States (US), 1 the majority of which will not meet their therapeutic goals despite treatment with multiple ADs. 2 Incretin-based drugs, including glucagon-like peptide 1 receptor agonists (GLP-1 RAs), are an important addition to the therapeutic options available for treatment of type 2 diabetes (T2D). 3 This class of drugs is associated with a reduction in glycated haemoglobin (HbA1c), weight loss and minimal risk of hypoglycaemia. [4][5][6][7] A reduced risk of cardiovascular events was also seen for some drugs in this class. 8,9 After introduction of the first GLP-1 RA, exenatide twice-daily, questions were raised about an increased risk of acute pancreatitis (AP) and pancreatic cancer (PC). The US Food and Drug Administration (FDA) and European Medicines Agency reviewed the totality of existing data but, to date, a final conclusion has not been rendered regarding a causal association between GLP-1 RA therapies and either AP or PC; therefore, they remain safety risks for these drugs. Both agencies called for more research in this area. 10,11 Thus, this post-marketing regulatory requirement supplements the growing body of studies utilizing real world data and methods to address concerns about confounding and bias in observational studies. A previously published brief report provided early results 12 ; this manuscript presents the final analyses.
Several studies using randomized clinical trial (RCT) methods or observational methods have already been published concerning this topic. Glycaemic control trials of liraglutide, 13 which belongs to the GLP-1 RA class, revealed more cases of AP in liraglutide arms as compared with control arms, although this was not confirmed in the Liraglutide Effect and Action in Diabetes: Evaluation of Cardiovascular Outcome Results (LEADER) trial. 14 While RCTs have been used in a number of evaluations, data from observational studies provide an important source of supplemental evidence. 15 Non-clinical data concerning the effects of GLP-1 mimetic drugs do not support a mechanism of action affecting the exocrine pancreas. 16 Numerous observational studies, as well as RCTs, have examined the risk of AP or PC in relation to GLP-1 RA and other incretin therapies (dipeptidyl peptidase-4 inhibitor [DPP-4i]), with mixed findings. 17 Recent meta-analyses, however, have provided increasing support that pancreatitis and PC are not a major concern in populations treated with incretin-based therapies. 15,18,19 This study aimed to estimate the risks of AP and PC in cohorts of patients initiating liraglutide as compared to initiators of other non-insulin ADs.

| Data source
This prospective cohort study was conducted using the Optum Research Database (ORD), which includes a large, geographically diverse population of commercial health insurance enrollees. The ORD contains enrollment data, medical claims and pharmacy claims; enrollment in 2014 included approximately 13.9 million individuals with both medical and pharmacy benefits and comprises approximately 4% of the US population. This study was approved by the New England Institutional Review Board and Privacy Board.

| Study design and study data
A new-user, active-comparator design was used for this study. 20 Patients were adult initiators of liraglutide or another non-insulin AD who had been continuously enrolled for at least 6 months prior to initiation of a study drug (baseline period), with no dispensing of that drug or drug class during baseline. for PC), whichever took place first. The AP cohorts were restricted to the sub-population for whom medical records could be requested to confirm AP outcome and patients with a diagnosis of AP or chronic pancreatitis (ICD-9577.1) at baseline were excluded. Patients with a PC diagnosis during baseline were excluded from both cohorts.

| Propensity Score (PS) matching
Logistic regression modeled the PS as the predicted probability of initiating liraglutide vs comparison treatment, conditional on covariates. 21 A number of covariates were included in the PS to address confounding, 22 with some forced into the model, including age, region, health care utilization metrics, number of unique ADs, diabetes severity, 23 AD use and outcome-specific risk factors. For AP cohorts, additional covariates included abdominal pain, cholelithiasis, cholecystectomy, pancreatic disease, overweight and smoking/alcohol use. 24 For PC cohorts, covariates for pancreatitis, other pancreatic disease, smoking/alcohol use and race were included. To account for changes in prescribing patterns and availability of ADs on the market, PS estimation and matching were performed within calendar periods for each comparator. 25 For the first three years, cohorts were created by calendar quarter and annually thereafter for the remaining recruitment time. Separately for AP and PC analyses, the eight comparator cohorts were matched 1:1 to liraglutide initiators within each period using a greedy matching algorithm. 26 Many patients initiated more than one study drug. Subsequent initiators who did not match during one recruitment time block, but who initiated any study drug during a later block, were again eligible for matching with PS, re-estimated using the baseline period prior to that initiation.
Patients could enter into multiple matched cohorts over the study period but were allowed to match into a drug cohort pair only once.

| Outcome definitions
Potential AP cases were initially identified by ICD-9 code 577.0, which does not include chronic pancreatitis, cysts or pseudocysts of the pancreas, or other diseases of the pancreas. Because the positive predictive value (PPV) for AP is typically low, 27 medical charts were reviewed by a gastroenterologist. The algorithm for confirming AP as "definite" or "probable" included: 3. Imaging diagnostic for, or suggestive of, AP. 28 Definite cases (three of the three criteria) and probable cases (criterion 1 and criterion 2 or 3) were considered confirmed cases. Final analyses included cases of AP that were confirmed through this medical chart review/adjudication. For each outcome, TOD analyses evaluated the effect of treatment within 'current', 'recent', and 'past' categories. "Current" use included first day of follow-up through end of days supplied, with an additional 31 days to account for medication non-adherence. Refill dispensings that were observed during these 31 days resulted in continuous "current" use time. "Recent" use began at the end of "current" time and continued for another 31 days. Subsequent person-time was categorized as "past" use, which persisted unless the patient reinitiated the same treatment, thereby re-entering "current" use.
Poisson regression was used to estimate the RRs and 95% CIs for "current," "recent" and "past" use of liraglutide vs the corresponding comparator category, adjusted by the logit of the PS to address confounding.
Another TOD analysis addressed potential for a dose-response relationship through calculation of cumulative time on liraglutide. Cumulative time exposed and unexposed to liraglutide during followup was quantified for PC, and IR was determined within each exposure type. Estimates of RR were made using Poisson regression within categories of cumulative time (short [<6 months], moderate  The results of AP analyses are presented in Figure 1. In the TOD analysis, RRs for "current" use ranged from 0.6 (95% CI, 0.2-1.8, sulfonylurea) to 1.9 (95% CI, 0.7-4.9, metformin). RRs varied substantially for "recent" use; however, these RRs were based on small numbers of cases and limited PYs, and thus could not be calculated for two comparisons. In general, RRs for "past" use were higher than estimates of "current" and "recent" use.
Most RRs were similar or lower when a one-year latency period was taken into account (data not shown).
In the TOD analysis, RRs for "current" use ranged from 0.    Diabetes-related retinopathy  In the ITT analyses, only the CI for all comparators except exenatide and DPP-4i excluded the null. Another sensitivity analysis that evaluated the potential selection bias that could result from preferential discontinuation following AP symptoms was completed for each of the eight matched AP cohorts. Baseline characteristics of patients who discontinued the study drug early during follow-up (≤median) and those of patients who discontinued use later during follow-up were compared. Occasional differences existed; for example, among patients who discontinued early, liraglutide patients were slightly older and more likely to be black than were DPP-4i patients. However, no consistent patterns of patient differences were observed overall, suggesting balanced F I G U R E 2 Propensity-score matched intention-to-treat and time-on-drug analyses for pancreatic cancer characteristics in the cohorts over time, regardless of timing in relation to discontinuation.
Several observational studies, including early results from this study, found no significant association between GLP-1 RAs (generally exenatide) and AP. 12,31-34 Garg and colleagues compared exenatide and sitagliptin initiators with a diabetic control group with participants initiating other ADs. The adjusted hazard ratio (aHR) for exenatide was 0.9 (95% CI, 0.6-1.5). Another cohort study that included liraglutide found no elevation in risk for "current" use of GLP-1 RAs  39 This result is in contrast to that of two similar meta-analyses of DPP-4i RCT that found significant increases in AP risk in the treatment group vs controls.
These studies were based on the same three RCTs with large patient populations, adjudicated cases of AP and median follow-up periods ranging from 1.6 to 3.0 years. 40 46 In addition, recent metaanalyses have provided additional evidence that PC is not a major concern in populations treated with incretin-based therapies. 19 The strengths of this study include the use of a large administrative claims database, supplemented with a medical record review.
Case ascertainment included medical record adjudication (AP) and a validated algorithm (PC). 29 Numerous ADs were evaluated, and cohorts were balanced in a wide range of potential confounders. An important limitation of studies sourced from health insurance claims data is that some important potential confounders (eg, BMI, smoking/alcohol use) may not be well characterized; however, PS matching has been shown to improve balance in unmeasured characteristics, including clinical parameters, as many measured covariates are proxies for unmeasured information. 22,47 Duration of follow-up may be limited in the ORD as the result of individuals changing health insurance plans; thus, the ability to assess the effect of study drugs on outcomes that occur more than 1 to 2 years after initiation is limited, which may affect the interpretation of results concerning those outcomes with longer induction periods (ie, PC).
Each cohort, including the all comparator cohorts, was created separately and was PS-matched with the liraglutide cohort, independently, within refined calendar periods to capture changes in US market share. Three all comparator cohorts were used, so that, in addition to including a matched comparison with all study comparators, one comparison involves non-GLP1-RA drug classes and another involves non-incretins.
Multiple analytic approaches were used to balance the strengths and weaknesses. Because AP is an acute event, TOD was the primary analysis, with a focus on "current" use. In contrast, longer-term exposure was important to assess the risk of PC; thus, ITT was the primary analysis, including latency analysis to reduce potential protopathic bias.
Medical record confirmation was required for AP cases, restricting the sample size to patients in plans that allow access to medical records (~30% of eligible patients). Despite this restriction, the number of AP cases is comparable to that of many other individual RCTs and observational studies with chart-confirmed outcomes. 19,39,48 The number of PC cases was small also because of the rarity of the outcome. PC cases were identified using an algorithm, possibly introducing misclassification; however, severity of PC, rapid time course once diagnosed, and use of an algorithm developed in the source database contribute to the accuracy of case identification. The number of cases was similar to that of those in studies included in pooled RCT analyses of GLP-1 drugs and in several other observational studies focused on PC. 19,43,44,46 Observational studies assess associations under "real-world" conditions; however, for a disease such as diabetes with numerous treatment options, these conditions complicate AD studies. Many patients have a history of AD use, and even restricted definitions of initiation may not fully capture earlier drug use. Concurrent or add-on drugs in other classes are difficult to quantify. Patients switch/add therapies if side effects occur or if diabetes is not well-managed. Baseline drug use was included in the PS; however, changes during follow-up are not accounted for with these analytic methods. As a result, it is difficult to attribute causality among patients using multiple drugs or undergoing multiple therapies.
Results should be interpreted with the understanding that comparators may not be reflective of initiators of those individual drugs in general, particularly for first-line therapies, e.g., the comparison is made between liraglutide initiators and metformin initiators with baseline characteristics similar to liraglutide initiators, and not a comparison to all types of metformin initiators.
Many measures and comparators were used to assess the association between liraglutide and pancreatic outcomes. AP analyses demonstrate no consistent pattern of increased risk with liraglutide during any period of use, including the primary analysis of "current" use.
There was no support for an increased risk of PC with liraglutide use, a finding that is consistent with many other studies.
In conclusion, based on observational data in a commercially insured population, liraglutide use was not associated with increased risk of AP or PC. Although high specificity of identified cases was achieved, concerns remain regarding the limited numbers because of the rarity of outcomes, as well as confounding by unmeasured factors and the question of whether the findings are generalizable to other populations.