Cardiac Risk Associated With Macrolides and Fluoroquinolones Decreases When Adjusting for Patient Characteristics and Comorbidities

Cardiac Risk and Shell Games—Finding the “Pea” (P-Value)

When I first read the title of the article by Polgreen et al, it piqued my interest; but, as I read on I was reminded of my first trip to a carnival. That metaphor comes to mind because this article has, like most carnivals, its own roller coaster and even a shell game. The reader is whipsawed back and forth by the authors’ statistical sleight of hand as we are given a glimpse of the pea (P-value) only to see it change. Their formidable use of statistical firepower is impressive; but, for this reader, also a confusing roller coaster ride. The authors analyzed 2 years of Medicare claims data for six cardiovascular events in post-MI patients to test the hypothesis that “after adjusting for important covariates, the studied antimicrobials will not have a clinically meaningful or statistically significant effect on the studied outcomes” (outcomes are identified by a dizzying array of 25 different ICD-9 codes). Of concern, this hypothesis and numerous statements in the article confuse causality with association, raising a question of objectivity and whether a statistical analysis of administrative claims data (a dull instrument at best) is an appropriate means to prove the absence of an association.

Any attempt by the average cardiologist to understand this statistical tour de force would likely require the reader to spend several hours with a biostatistics consultant. So, as readers, we assume that the article’s reviewers have carefully evaluated the statistical methods and the analysis can be trusted, even if not fully understood. Data from 185,000 patients aged ≥66 years and discharged after acute MI were analyzed for cardiac and vascular events in the week after Medicare Part D was billed for one of the antibiotics of interest (azithromycin, clarithromycin, levofloxacin, moxifloxacin, amoxicillin, or doxycycline). Standard logistic regression analysis was performed and then followed by adjustments for covariates, and multiple comparisons were made to determine if the increased cardiac risk associated with antibiotics might be confounded by indication. The authors asked, “Is simply having an infection or some factor(s), other than the antibiotic, associated with an increased cardiovascular risk?” The results are complex but interesting.

Like a pea in the shell game, the P-values jump from extreme to extreme as the analysis plays out, each one contradicting the others. Initially, the analysis suggests that all antibiotics they studied are associated with increased cardiac events (a finding in numerous prior studies except for doxycycline). Next, the authors present an “adjusted” analysis and all the “peas” disappear as the odds ratios (OR) for any of the six cardiac outcomes fell from 1.35 to 1.01. This adjustment of OR involved the consideration of >108 covariates (selected from 143; not all listed in the paper). These included some that are reasonable but most are without any obvious justification, such as English-speaking percentage and low-income subsidy. Recognizing the risk of making 42 comparisons, the authors applied the Bonferroni correction, which required them to use P-values of ≤0.0012 for significance. The adjusted ORs led the authors to conclude, “According to the adjusted analyses, none of the 6 studied antimicrobials significantly increased the risk of death.” Here again, the authors may have displayed bias because this remarkable conclusion invokes causality instead of their failure to observe “associations” between drugs and events.

With their hypothesis considered proven, the authors delve further into the data and speculate that some of the drugs are perhaps lifesaving, and state that “the risk of death is significantly decreased for azithromycin and levofloxacin.” Again, the author’s choice of words suggests drug causality when they are, in fact, merely identifying associations. Amazingly, the last stop on the roller coaster occurs when the authors backtrack and state, “We cannot dismiss the possibility that these antimicrobials are associated with some increase in cardiovascular risk.” As humble readers we ask which of these four dramatically different conclusions should we accept? Where is the pea to be found?

The authors dutifully disclose the fact that their study is subject to several limitations such as the fact that the cohort is limited to include only post-MI patients and that only those antibiotics paid for by Medicare Part D would appear in their data. One cannot assess the potential impact of any drugs paid for by secondary insurers or by patients out of pocket. Other weaknesses are mentioned, such as the large number of multiple comparisons (n = 42) and covariates (n = 135). However, many others are not mentioned, such as the well-known limitations of using ICD-9 codes.

What does this paper tell us that we didn’t already know? Extensive prior research has shown that antibiotics are extremely safe but, in some populations, macrolides and fluoroquinolones have been associated with increased mortality and cardiac events. This article’s complex analysis of claims data for post-MI patients argues that increased cardiac risk is confounded by the illness that prompted the prescription of an antibiotic. Yet, the many other studies, which include prospective randomized trials where such confounding should not be a factor, are not referenced or discussed.

Importantly, the article fails to address the potential role of the unquestioned cardiac toxicity of macrolides and fluoroquinolones—that is, torsades de pointes (TdP)—and the fact that the diagnosis of this potentially lethal arrhythmia is often missed by clinicians and billing coders. Importantly, at the time the data were gathered for this study, TdP had no specific ICD-9 code, making it necessary for the authors to use imprecise surrogate ICD-9 codes.

No one should question the fact that TdP, although rare, is a very real consequence of administering macrolides and fluoroquinolones to high-risk patients. It is difficult to detect TdP in large administrative databases, especially when the analysis is limited to those patients with a single, minor risk factor like prior MI. TdP is most likely to be detected in those who are elderly, female, and patients with heart disease, hypokalemia, bradycardia, and/or concomitant drugs that prolong repolarization. This particular form of cardiac toxicity should not be confused with other forms of cardiac risk because their etiology and treatment are quite different.

For this reader, the take-home message from this article is that any cardiac risk with antibiotics is likely to be very small, especially in a general population, and should not prevent any patient with a valid medical indication from being treated with any of these antibiotics. The corollary is that antibiotics, like all drugs, have some known risks and should not be prescribed inappropriately, such as for viral infections. Lastly, we must be cognizant of the potential for macrolides and fluoroquinolones to prolong QT and to result in TdP in at-risk patients. Anticipating and managing this risk requires awareness by healthcare providers and the use of decision-support systems such as those now being employed in many major medical centers. As readers and reviewers of the medical literature, we must insist that scientific articles present a consistent interpretation of scientific findings that are supported by the most rigorous interpretation of the data. I find this article to be lacking in consistency of data interpretation and objectivity in reaching conclusions.