Adverse events associated with incretin-based drugs in Japanese spontaneous reports: a mixed effects logistic regression model

Data source

The JADER dataset, which was published in July 2015 and contained 353,988 unique cases, was obtained from the website of the PMDA. The analyzed cases were reported from January 2010 to March 2015 and had available records regarding age and sex.

Adverse events in the JADER were coded as “Preferred Terms” (PTs) in the Japanese version of the Medical Dictionary for Regulatory Activities (MedDRA/J) (MedDRA, 2015). MedDRA has a hierarchical structure, in which PTs are grouped into “High Level Terms” (HLTs), HLTs are grouped into “High Level Group Terms” (HLGTs), and HLGTs are grouped into “System Organ Classes” (SOCs). Before data analyses, a relational database was constructed from the JADER dataset and MedDRA/J version 18.0. SQLite version 3.8.5 was used as the database management system (SQLite Development Team, 2015).

As incretin-based drugs, all DPP-4 inhibitors and GLP-1 receptor agonists approved in Japan were assessed. The DPP-4 inhibitors were sitagliptin phosphate hydrate, vildagliptin, alogliptin benzoate, alogliptin benzoate/pioglitazone hydrochloride (combination drug), linagliptin, teneligliptin hydrobromide hydrate, anagliptin, and saxagliptin hydrate. The GLP-1 receptor agonists were exenatide, liraglutide, and lixisenatide.

Data analysis

The analysis of ADEs associated with incretin-based drugs was composed of two phases. The first phase was a DPA based on Fisher’s exact test. The second phase was a multivariate analysis using a mixed effects logistic regression model.

The PTs of ADEs were classified according to the MedDRA HLTs. All combinations of generic drug names and HLTs were extracted. Fisher’s exact tests were performed for all combinations of incretin-based drugs and reported HLTs. Associations that yielded a two-sided p-value <0.01 and an OR >1 were considered significant.

Mixed effects logistic regressions were performed for each HLT significantly associated with incretin-based drugs. The mixed effects logistic regression model was as follows: $\frac{P\left(Y_i=1|x_i,z_i\right)}{P\left(Y_i=0|x_i,z_i\right)}=exp\left(x_i^T\beta +z_i^{T}u\right)$ Where Y_i is a binary variable describing the outcome of case i (0 or 1), β is a fixed parameter vector, x_i is a covariate vector for fixed effects, u is a vector of random variables from probability distributions, and z_i is a covariate vector for random effects. u represents unmeasured covariates as a way of modeling heterogeneity and correlated data (Larsen et al., 2000).

In the newly developed model, the binary outcome was whether or not each HLT was reported. For fixed effects, the covariates were use of DPP-4 inhibitors, use of GLP-1 receptor agonists, use of any hypoglycemic drugs (an alternative indicator for hyperglycemia), sum of concomitant suspected drugs (determined by reference to the Fisher’s exact tests), age (in 10-year intervals), and sex. The random effect was the quarterly period of reporting. The variables for the random effect were random intercepts normally distributed with mean 0 and one common variance. The associations between incretin-based drugs and HLTs were assessed by ORs with 99% two-sided Wald-type Confidence Intervals (CIs). Because the present analysis was an exploratory screening, the problem of multiple comparison was not addressed. Instead of correcting that, stringent levels of statistical significance were set (p-value <0.01 and 99% CI).

The newly developed mixed model was compared with a fixed model that did not include the random effect. The covariates for fixed effects in the fixed model were the same covariates use in the mixed model. Logistic regressions based on each model were performed for all reported HLTs associated with incretin-based drugs. Subsequently, the adequacy of the model was assessed by Akaike’s Information Criteria (AIC) (Burnham & Anderson, 2002). A model with a smaller AIC was favored.

All analyses were performed using the R version 3.2.1 (R Development Core Team, 2010). The glmmML package version 1.0 was used with the “ghq” (Gauss-Hermite quadrature) method for the mixed effects logistic regressions (Broström, 2013).

Description of the analyzed case reports

The JADER included 204,472 unique cases that were reported from January 2010 to March 2015, of which 187,181 had available records for age and sex and were analyzed. The records included 4,952 generic drug names and 6,151 PTs under 1,377 HLTs. DPP-4 inhibitors were mentioned in 7,265 cases, whereas GLP-1 receptor agonists were mentioned in 451 cases. Figure 1 shows the number of cases mentioning hypoglycemic drugs that were reported during each quarterly period. Although the number of cases for other hypoglycemic drugs increased gradually over time, the number of cases for DPP-4 inhibitors increased markedly.

Mixed effects logistic regressions

The cases associated with incretin-based drugs included 1,430 PTs under 735 HLTs. The Fisher’s exact tests showed that 106 of the 735 HLTs were significantly associated with any incretin-based drug (two-sided p-value <0.01 and OR >1). In the mixed effects logistic regressions, 33 of the 106 HLTs identified by the Fisher’s exact tests were significantly associated with DPP-4 inhibitors or GLP-1 receptor agonists (99% two-sided CI). Table 1 shows the number of cases reported for each HLT. Figure 2 shows ORs with 99% CIs for the significant associations between HLTs and DPP-4 inhibitors or GLP-1 receptor agonists. “NEC” in the MedDRA terms is an acronym for “Not Elsewhere Classified,” which denotes groupings of miscellaneous terms, whereas “excl” is an abbreviation of “excluding.” The HLTs associated with DPP-4 inhibitors included “Pancreatic disorders NEC” (OR 18.66; 99% CI 2.09–166.25) and “Acute and chronic pancreatitis” (8.65; 5.76–12.98). The HLTs associated with GLP-1 receptor agonists included “Thyroid neoplasms” (87.25; 6.64–1146.27) and “Cystic pancreatic disorders” (61.32; 1.69–2224.49). The HLTs associated with DPP-4 inhibitors and GLP-1 receptor agonists indicated pancreatic events (“Acute and chronic pancreatitis,” “Pancreatic neoplasms,” “Pancreatic neoplasms malignant (excl islet cell and carcinoid),” and “Pancreatic disorders NEC”), gastrointestinal events (“Benign neoplasms gastrointestinal (excl oral cavity)” and “Gastrointestinal stenosis and obstruction NEC”), and cholecystic events (“Cholecystitis and cholelithiasis”). Although DPP-4 inhibitors and GLP-1 receptor agonists were not associated with hypoglycemic events, GLP-1 receptor agonists were associated with several HLTs related to diabetes, including “Hyperglycaemic conditions NEC” and “Diabetic complications NEC.”

Comparison between the models with and without the random effect

Figure 3 shows the comparison between the models with (mixed model) and without (fixed model) the random effect. In 604 of the 735 HLTs reported with incretin-based drugs, the AIC of the models were calculated normally. Of the 604 HLTs, 302 favored the mixed model, whereas the others favored the fixed model. The median number of reported cases for the group of HLTs favoring the mixed model was 264, whereas the median number of reported cases for the group of HLTs favoring the fixed model was 83; thus, HLTs reported frequently tended to favor the mixed model.

Time-series variation of spontaneous reports

SRSs accumulate a large amount of data regarding ADEs every year; thus, the contents of SRSs are not constant. In the present study, the report composition of hypoglycemic drug groups varied during the study period. Reports associated with DPP-4 inhibitors showed a marked increase in comparison with that of reports associated with other hypoglycemic drugs, perhaps because of an increasing number of approved products and an associated increase in drug use. The number of reports during a particular period is affected by numerous factors. This variation in reporting results in temporal heterogeneity, supporting the appropriateness of the mixed model.

Adverse events associated with incretin-based drugs

Some HLTs associated with incretin-based drugs in the present study have been reported as issues of concern in previous studies. Some groups of similar HLTs, e.g., “Thyroid neoplasms” and “Thyroid neoplasms malignant” were identified because some PTs are linked to multiple HLTs in the MedDRA. In comparison with DPP-4 inhibitors, GLP-1 receptor agonists showed relatively wide CIs for some of the HLTs because fewer cases were reported, leading to unreliable results.

In the present study, pancreatic disorders, including pancreatitis and pancreatic cancer, were associated with DPP-4 inhibitors and GLP-1 receptor agonists, which were consistent with results obtained via analysis of the FAERS (Butler et al., 2013; Elashoff et al., 2011). In addition, thyroid cancer was associated with GLP-1 receptor agonists; however, because of the small number of cases, this finding is unreliable. Analyses of the FAERS also indicated that GLP-1 receptor agonists increased ORs for thyroid cancer (Butler et al., 2013; Elashoff et al., 2011). Thyroid cancer and pancreatic disorders are among the most controversial safety concerns regarding incretin-based drugs; however, no evidence has been found for such risks in human clinical studies (Butler et al., 2013; Nauck, 2013). The other HLTs associated with DPP-4 inhibitors and GLP-1 receptor agonists were “Benign neoplasms gastrointestinal (excl oral cavity),” “Gastrointestinal stenosis and obstruction NEC,” and “Cholecystitis and cholelithiasis.” Gastrointestinal events such as nausea, vomiting, and diarrhea are common ADEs of incretin-based drugs (Nauck, 2011); however, benign gastrointestinal neoplasms, stenosis, and obstruction have not been referred to in past studies. Cholecystic events have not in the same way.

Hypoglycemia, an adverse event associated with some hypoglycemic drugs, was not associated with incretin-based drugs. In contrast, hyperglycemia and several other diabetic complications were associated with GLP-1 receptor agonists, perhaps because of cases of ineffective drug treatment.

Limitations

The limitations of SRS data mining include confounding by indication (i.e., patients taking a particular drug may have a disease that is itself associated with a higher incidence of the adverse event), systematic under-reporting, questionable representativeness of patients, effects of media publicity on numbers of reports, extreme duplication of reports, and attribution of the event to a single drug when patients may be exposed to multiple drugs (Gibbons et al., 2010). In addition, spontaneous reports do not reliably detect adverse drug reactions that occur widely separated in time from the original use of the drug (Brewer & Colditz, 1999).

The newly developed model reported here addresses the confounding influences of temporal heterogeneity and concomitant drug use. Nevertheless, risks identified via analysis of SRS data should be considered as safety signals, rather than definitive statements of cause and effect. For further interpretation of each ADE, additional reviews of other data sources are recommended.

Mixed effects logistic regression model

In the AIC comparison between the mixed model and the fixed model, half of the HLTs reported with incretin-based drugs favored each model. The HLTs that favored the mixed model were reported more frequently than those that favored the fixed model, indicating that the mixed model may be more appropriate for frequently reported ADEs. The AIC formula has a bias-correction term for the number of estimable parameters (Burnham & Anderson, 2002). In the above comparison, the mixed model has only one more parameter than does the fixed model; hence, the difference between the penalties for the correction is small.

The adequacy of the random effect was demonstrated; however, the model can be improved. Although we assumed a normal distribution for the random effect, the appropriateness of this assumption is unclear. Moreover, it is unclear whether a single probability distribution is sufficient to assess the random effect on the widely spread time-scale of spontaneous reports. Sampling of parameter distributions by Bayesian hierarchical modeling is a potential solution to these problems. Currently, diverse implementations of Bayesian methods, which support practice of such modeling, are accessible (Li et al., 2011; MacLehose & Hamra, 2014).