Author + information
- aDepartment of Medicine, Vanderbilt University Medical Center, Nashville, Tennessee
- bDepartment of Pharmacology, Vanderbilt University Medical Center, Nashville, Tennessee
- cDepartment of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, Tennessee
- ↵∗Reprint requests and correspondence:
Dr. Dan M. Roden, Vanderbilt University Medical Center, 2215B Garland Avenue, 1285 MRBIV, Nashville, Tennessee 37232-0575.
Rare serious adverse drug reactions (ADRs) continue to present an important problem both for clinicians and their patients as well as for the drug development community. In the clinical arena, we would like to know which drugs are likely to cause an ADR and in which patients (1), whereas the drug developers (2) need to understand which candidate molecules have a high liability for producing ADRs. Although a tremendous amount of work at the clinical, molecular, and genetic levels has defined important mechanisms contributing to drug-induced QT prolongation and risk for torsades de pointes (the drug-induced long QT syndrome [diLQTS]), the datasets remain incomplete. Clinical studies have identified factors increasing risk for diLQTS, including female sex, baseline QT prolongation, unrecognized congenital long QT syndrome, other genetic variants, hypokalemia, and bradyarrhythmia (3,4). In addition, we think we understand how to identify high-risk drug candidate molecules by screening for liability to block IKr (a major repolarizing current in heart carried by the Kv11.1 channel, encoded by HERG, more formally known as KCNH2) or, as has recently been suggested, to increase late sodium current (4). Action potential prolongation in vitro and QT prolongation in human subjects are the hallmarks of a drug at risk.
Studying ADRs due to single drugs is tough; however, with a demographic shift toward an older population in whom polypharmacy is the norm, ADRs due to drug interactions adds even greater complexity to this problem. In this issue of the Journal (5), the Tatonetti laboratory (6) at Columbia University has made an important new contribution which will only further complicate decision making by clinicians and by regulators. Using a three-step combination of data mining approaches and electrophysiological methods in vivo and in silico (Figure 1, left), Lorberbaum et al. (5) have defined a combination of commonly used drugs (ceftriaxone + lansoprazole) that, individually, have no QT “signal,” but in combination appear to increase liability for diLQTS.
What did they Find? How Strong is their Evidence?
The group started with an evaluation of the U.S. Food and Drug Administration’s Adverse Event Reporting System (FAERS). They implemented an interesting data mining approach that does not search directly for events like QT prolongation or arrhythmia (6,7) but rather first identifies from electronic health records (EHRs) common clinical features associated with the use of drugs that do or do not prolong the QT interval. This is similar to the approach previously used to identify a drug-drug interaction predisposing to glucose intolerance: rather than searching directly for diabetes, the algorithm searched for patients who “looked like” they had diabetes (8). Using these clinical features as discriminators, the authors searched in FAERS for a series of target-related ADRs such as “ECG QT prolonged” and “torsades de pointes” (e.g., “arrhythmia” or “atrial fibrillation”), and unrelated ADRs (“rhabdomyolysis” or “complete suicide”). Furthermore, rather than searching for drugs reported to produce these effects, they sought pairs that individually did not appear to produce arrhythmia signals but did in combination. This initial mining approach identified 889 potential interactions associated with QT prolongation. The messiness of the FAERS dataset is illustrated by the fact that some of these combinations, including the ceftriaxone/lansoprazole combination on which the remainder of the study focuses, modestly increased risk for other effects such as rhabdomyolysis, agitation, and completed suicide in the initial dataset. Pharmaceutical sponsors of new drugs are obligated to report to FAERS events, in clinical trials or even described as anecdotes to members of their sales forces. Reporting by healthcare professionals is allowed but not compulsory, and reporting is also allowed by the public. We do not have a sense beyond the manufacturers for what motivates those who report events to FAERS and whether they are at all representative of what goes on in practice.
The second step was to examine these 889 potential signals in the Columbia EHRs. Thirty-four cases passed an initial evaluation, but of these, 26 records included drugs known to prolong QT interval individually. The investigators then focused on 1 of the remaining 8, the combination of ceftriaxone and lansoprazole, given the frequency with which these 2 drugs are co-administered. The report states that QT intervals for male patients “taking this combination” were 12 ms longer than for male patients taking either drug alone. This is actually an overstatement because the methods used searched for co-administration required at least 1 dose of each of the 2 drugs within a 7-day window and an electrocardiogram (ECG) recorded up to 36 days after this potential co-administration. Therefore, there is a possibility that the drugs were not actually co-administered and that the ECGs were not actually recorded during any co-administration. However, potential misclassifications in the EHR like these typically result in bias toward a null result. One can also ask what co-morbidities were present (e.g., extent of heart disease) that might have potentiated or even caused QT prolongation observed under those clinical conditions. Indeed, the numbers of subjects with at least 1 ECG reporting a QTc value >500 ms was astoundingly high (19% in the combination group and 14% of subjects taking 1 of the 2 drugs), suggesting that the rate of co-morbidities might have been quite high.
The third step was to evaluate the effect of the 2 drugs alone and in combination on IKr in vitro. This approach makes the assumption that the interaction is pharmacodynamic (i.e., the 2 drugs combine somehow to increase IKr block), rather than pharmacokinetic, whereby the effects of the combination arise from interaction-based changes in drug concentrations. In this case, there is little evidence in support of a pharmacokinetic mechanism, although other interactions, such as that between quetiapine and methadone (9) as discussed in the article (and that might even be detectible using the approaches adopted here), probably do have a pharmacokinetic basis. Ceftriaxone alone produced no effect on IKr, and lansoprazole (10 μM) reduced IKr ∼14%. The combination of ceftriaxone with a very low lansoprazole concentration (0.1 μM) reduced IKr by ∼37%; interestingly, a lansoprazole concentration 1,000-fold higher (100 μM) produced only a small additional effect (−58%). This is an unusual dose-response curve and suggests some mechanism beyond simple additivity. The electrophysiological experiments also included a combination not found to produce diLQTS in the EHR analysis (cefuroxime + lansoprazole), although the number of subjects potentially exposed to this combination was much less than those potentially exposed to ceftriaxone + lansoprazole. The interaction was, interestingly, slightly more prominent in males than in females and appeared slightly more prominent in subjects of European rather than African ancestry. The authors point out that the frequency of common potassium channel polymorphisms may vary by ancestry. This is the case for virtually all genes and, in the case of ion channels, has been used to invoke ancestry-specific arrhythmia susceptibility (10,11).
Putting Together Multiple Datasets
As briefly outlined above, each set of evidence contains important flaws and leaves open methodological questions. However, each set of evidence provides support for the underlying hypothesis that this drug interaction does, in fact, unexpectedly prolong the QT interval. The results and their interpretation provide important lessons for investigators interested in using “Big Data” approaches to study ADRs, other drug effects, and indeed, many other aspects of the human condition.
In the domain of genome science, for example, there has been a large and well-justified focus on generating very large datasets for replication of initial signals. This approach has its greatest utility when applied to common phenotypes that can be ascertained across hundreds of thousands of subjects. For rarer phenotypes or those more difficult to define, such as those estimating response to drugs and, in particular, serious ADRs, the numbers of subjects accrued in any study will necessarily be much smaller. How then, can a community ever derive confidence that a given result generated across few subjects is real? The present study provides one answer: by considering multiple layers of evidence from different (“orthogonal”) datasets, investigators (and readers of JACC) can derive increased confidence that a given result is, in fact, “real,” as seems to be the case here (Figure 1, right). Is this result robust enough to advise clinicians to avoid this drug combination in all patients or in patients at risk for QT prolongation? Probably not at this point. What would really add value would be a well-controlled study (which need not be large) in human subjects, examining the effects of the drugs individually and in combination on the QT interval. That would be yet another orthogonal piece of data and would provide very strong evidence in support of the plausible hypothesis that ceftriaxone and lansoprazole together increase risk for diLQTS, and would quantify that combined effect. A major ongoing challenge to the approach of combining diverse datasets is to develop appropriate statistical methods to critically evaluate, beyond a “gut feeling,” that a given orthogonal set of findings is real. With polypharmacy becoming the norm, the number of potentially interacting drug pairs is vast, and testing all possible combinations for ADRs in the relevant populations is not feasible but will require instead the routine surveillance of larger and larger real-world data sets. Hence, solving the methodological challenges of developing approaches to systematically leverage these data sources will be the next frontier in identifying and preventing ADRs.
↵∗ Editorials published in the Journal of the American College of Cardiology reflect the views of the authors and do not necessarily represent the views of JACC or the American College of Cardiology.
Supported by U.S. National Institutes of Health grants GM115305 and HL49989. The authors have reported that they have no relationships relevant to the contents of this paper to disclose.
- American College of Cardiology Foundation
- Blaus A.,
- Madabushi R.,
- Pacanowski M.,
- et al.
- Kaab S.,
- Crawford D.C.,
- Sinner M.F.,
- et al.
- Roden D.M.
- Lorberbaum T.,
- Sampson K.J.,
- Chang J.B.,
- et al.
- Tatonetti N.P.,
- Fernald G.H.,
- Altman R.B.
- Tatonetti N.P.,
- Ye P.P.,
- Daneshjou R.,
- Altman R.B.
- Splawski I.,
- Timothy K.W.,
- Tateyama M.,
- et al.
- Hwang H.W.,
- Chen J.J.,
- Lin Y.J.,
- et al.