Use of an algorithm for identifying hidden drug–drug interactions in adverse event reports

In a recent JAMIA article, Tatonetti et al¹ employed a novel algorithm to identify drug–drug interactions in spontaneous reporting systems, for example, the US Food and Drug Administration adverse event reporting system. The authors present two examples of drug–drug interactions identified by the algorithm: (1) moxifloxacin and warfarin and increased renal impairment, and (2) paroxetine and pravastatin and diabetes-related events.

Tatonetti et al¹ performed a retrospective database study of electronic medical record (EMR) data to evaluate the interaction between moxifloxacin and warfarin and the increased risk of renal impairment. An analysis of covariance was conducted to test for differences in risk for patients taking moxifloxacin and warfarin compared to patients taking other antithrombotic agents and fluoroquinolones. Crude results showed an increase in risk of renal impairment among those on moxifloxacin and warfarin; however, no association was detected after adjusting for baseline covariates.

A retrospective database study of EMR data was also conducted by Tatonetti et al² to evaluate the paroxetine and pravastatin interaction and diabetes-related adverse events, specifically an increase in random blood glucose measurements. Validation of the paroxetine and pravastatin interaction was completed via analyses of EMR data from three sites: Stanford University Hospital, Vanderbilt Hospital, and Partners HealthCare. Details of the validation methods are published elsewhere.² Random glucose measurements were extracted before and after treatment with paroxetine and pravastatin, alone and in combination. An analysis of covariance was conducted to evaluate changes in blood glucose levels from baseline. The authors stated that a ‘traditional covariate analysis’ was not possible for the study.² They concluded that paroxetine and pravastatin were associated with a significant increase in blood glucose in patients taking both drugs. However, the clinical significance of these findings is unknown.

We applaud the authors’ efforts to develop a novel method for the identification of potential drug–drug interactions. However, like the moxifloxacin and warfarin interaction, we question whether the perceived risk due to the combination use of paroxetine and pravastatin could also be explained through statistical adjustment of baseline comorbidities. The perceived risk could also be a result of selection biases such as confounding by indication. In addition, we question whether diabetes-related adverse events, including increases in non-fasting blood glucose measurements, actually lead to a clinically significant event, that is, an increased risk of developing type 2 diabetes mellitus (T2DM).

To evaluate this potential signal further, we are conducting a retrospective cohort study using large claims databases, OptumInsight and Thomson Reuters Marketscan. All patients 18 years of age and older who were newly prescribed paroxetine or pravastatin are included in the study. Patients were categorized as users of paroxetine alone if there were no existing prescriptions for statins during the 6-month baseline period or during follow-up. Likewise, patients were categorized as users of pravastatin alone if there were no existing prescriptions for selective serotonin reuptake inhibitors during the baseline or follow-up periods. Combination therapy is defined by a new prescription of paroxetine within 45 days of an existing pravastatin prescription, or a new prescription of pravastatin within 45 days of an existing paroxetine prescription.

Preliminary results of OptumInsight data show no association between the concomitant use of paroxetine and pravastatin and new onset T2DM compared to paroxetine alone (HR 1.09; 95% CI 0.60 to 2.01) or compared to pravastatin alone (HR 1.05; 95% CI 0.77 to 1.45). Results are adjusted for age, gender, region, year of cohort entry, comorbid conditions, and co-medications during the baseline period. A detailed description of the methodology and results will be published at the study's conclusion. Our findings directly contradict the conclusions of Tatonetti et al² about the clinical importance of increases in blood glucose levels in users of paroxetine and pravastatin.

Our preliminary analysis examined the association between the drug–drug interaction and the clinical outcome of new-onset T2DM. Although the authors chose to examine changes in glucose measurements, we believe the examination of unstable laboratory values, such as non-fasting glucose measurements, can result in misleading conclusions. It should also be noted that laboratory test results in observational data are not available for all patients uniformly; values are likely to be available only for patients with conditions that could confound perceived associations. Therefore, statistical adjustment is imperative.

The authors state that their new method addresses the issue of underreporting and could have the ability to identify ‘hundreds of novel interactions’. Nevertheless, signals produced by the authors’ method, like all signal detection methods, must be evaluated and confirmed using robust epidemiological and statistical methods and clinically important endpoints. Whether this new method discovers more signals with a better positive predictive value than current methods remains to be confirmed.