We reported an effect of proton pump inhibitors (PPIs) leading to vascular dysfunction (Circulation 2013;128:845–853). Based on that finding, we hypothesized that PPIs may be associated with an increased risk of myocardial infarction, and reapplied a validated data-mining method to examine for the presence of a drug safety signal in electronic medical records. With regard to the critique of our paper (PLoS One 2015;10:e0124653), the commentators have attributed many of the limitations of observational studies to our data-mining study. However, we believe that they fail to account for the unique merits of data mining methodology as a tool to identify safety signals with medications and to advance pharmacovigilance. First, let us review the difference between observational studies and data-mining studies. Data-mining studies (or algorithmic modeling as described by Breiman in Statistical Science 2001;16:199–231) treat the causal mechanism as an unknown and attempt to learn a function f(x) that operates on variables (x) to predict the responses (y). Therefore, the linking function f(x) in a data-mining study should not be interpreted as a causal regression model. The commentators have critiqued our work not appreciating the distinction between the two statistical approaches. The validation of data-mining approaches is performed by measuring predictive accuracy using a set of known true and false associations. In performing such validation of data-mining studies—using a set of known associations, rather than “adjusting” for specific hypothesized confounders on a case-by-case basis—the key idea is that if confounding is present, upon testing known negative associations, those will signal as positive (or vice versa a positive association will be missed). A poorly designed data-mining process will thus make the wrong call more often than a well-designed process. This approach, of estimating how likely are we to be wrong, thus depends on the breadth of the set of true and false associations used. In 2013, we validated our data-mining pipeline using a set of 28 true-positive and 165 negative associations (specificity of 97.5%, sensitivity of 39%, with an area under the curve of 80.4%; Clin Pharmacol Ther 2013;93:547–555), and earned an editorial describing the work as representative of the cutting edge of drug safety and pharmacovigilance science (Clin Pharmacol Ther 2013;93:474–475). As described in (Clin Pharmacol Ther 2013;93:547–555) and reproduced in Supplemental Figure 1 for the PLoS article, temporality is addressed. In fact, some of the situations speculated, such as confounding owing to PPI use as first-line stress ulcer prophylaxis are avoided by the flowchart described in Supplemental Figure 1. Finally, we also assessed the issue of outcome validity by evaluating 25 outcomes by manual review in our 2013 paper (Supplemental Figure 3; Clin Pharmacol Ther 2013;93:547–555). Second, the possible mechanism that we proposed has been experimentally tested—both in vitro and in vivo, with validation by a second investigative group using a different experimental approach to assess vascular function—and published in 2013 (Circulation 2013;128:845–853). In addition, there are several studies that raise concerns about the use of PPIs in patients with coronary artery disease, as well as in the general population (Ann Intern Med 2010;153:378–386; BMJ 2011;342:d2690; Circulation 2012;125:978–986; JAMA Intern Med 2013;173:518–523). Shih et al (Int J Cardiol 2014;177:292–297) examined this association in the general population, reporting a 1.58-fold greater risk of myocardial infarction with PPI use. Short-term use of PPIs probably has little adverse cardiovascular effect. Indeed, in a small, randomized trial in healthy volunteers, the short-term exposure to PPIs did not have a significant effect on plasma asymmetric dimethylarginine and vascular function (Vasc Med 2015;20:309–316). Overall, the cardiovascular risk from PPI use in the general population remains poorly studied. Therefore, given our findings that PPIs may adversely affect vascular function, given the association seen via data-mining in two unrelated datasets, given the association seen on reanalysis of a prospective dataset, and given a separate independent study (Int J Cardiol 2014;177:292–297), the issue of cardiovascular risk from PPI use in the general population needs further investigation. We did not conclude with a recommendation to change practice based on our findings; in fact, we explicitly say that the work presented provides provide an example of how a combination of experimental studies and data-mining approaches can prioritize drug safety signals for further investigation. Given that PPIs were approved by the US Food and Drug Administration for short-term use (weeks, rather than months or years) and the current standard of practice deviates from that, if further research confirms the risk, physicians and patients should assesses the risks and benefits of long-term PPI use. Is Proton Pump Inhibitor Use Associated With Risk of Myocardial Infarction?GastroenterologyVol. 150Issue 2PreviewShah NH, LePendu P, Bauer-Mehren A, et al. Proton pump inhibitor usage and the risk of myocardial infarction in the general population. PLoS One 2015;10:e0124653. Full-Text PDF
Read full abstract
Apr 19, 2017
Yvonne Y Lau + 6
Open Access