A vital component of patient safety is screening for drug-drug interactions. As I described in my last post, this type of screening is one of the requirements to achieve meaningful use of electronic health records.

Drug Interactions
A drug-drug interaction occurs when the action of one drug is influenced by the action of another drug. While drug interactions can sometimes be beneficial, in safety screening we are most concerned with drug interactions that have the potential to cause harm. Even within this group, some interactions are more clinically significant than others. In some cases it is clear from human studies that using two drugs together puts the patient at an unacceptable risk of harm. In many instances, the two drugs may be safely administered together, but additional patient monitoring may be appropriate to detect possible adverse drug effects. In other cases the studies may have been inconclusive, and the effects, even if real, might be minor.

If we consider clinically significant interactions to be signal and other, minor interactions to be noise, we want to maximize the signal:noise ratio in any computer software we deploy for drug-drug interaction screening. If most of the interactions displayed are minor and wouldn’t necessitate a change in a prescription, providers start to ignore the alerts. Eventually, they may ignore all of the alerts, even those that are significant.

JAMIA study
The ability for drug-drug interaction software to detect clinically significant interactions is therefore extremely important. In the January 2011 issue of JAMIA, Saverno et al studied the ability of the software in use at 64 pharmacies in Arizona to detect clinically significant interactions. At each pharmacy, the researchers entered a fictitious patient with 18 prescriptions. Of the 153 possible interaction pairs, the authors pre-identified 13 drug-drug interactions as being clinically significant.

The authors also identified 6 pairs as non-interacting. They did not classify the other 134 possible drug-drug pairs. Across all the pharmacies in the study, the median sensitivity was 0.85, meaning that the pharmacy software detected 85% (about 11) of the 13 drug-drug interactions that the authors deemed to be clinically significant. Some pharmacies had a sensitivity lower than the median (as low as 0.23), while others did better (as high as 1.0). The median specificity was 1.0, meaning that the pharmacy software correctly did not report the 6 non-interacting pairs as interactions.

Medi-Span
I took the opportunity to enter this fictitious patient profile into our Medi-Span Clinical product, which has drug interaction screening powered by the same clinical data as all of Medi-Span’s drug interaction products. Medi-Span Clinical identified all 13 clinically significant interactions and did not identify any of the 6 non-interacting pairs. Using the author’s gold standard, Medi-Span Clinical therefore had 100% sensitivity and 100% specificity. As I noted above, the authors didn’t classify the other 134 possible pairs, so we can only compute sensitivity and specificity as the authors did in the study, using the 19 pairs that they classified.

Medi-Span Clinical also identified several other interactions that I would consider to be clinically significant, such as an interaction between fluconazole and amiodarone, which combination can prolong a part of the cardiac cycle (the “QT interval”) and increase the risk of a potentially fatal cardiac arrhythmia (“Torsades de Pointes”).

Current State
The study demonstrates that as an industry, we certainly have room to improve when it comes to identifying clinically significant interactions and reducing noise. We don’t know which pharmacies in the study were using which drug databases, but we know that some pharmacies had excellent results. We also know that Medi-Span Clinical provided excellent results in my own analysis. Therefore, in contrast to the median state, in which systems can miss 1 in 7 clinically significant interactions, we can describe the present state-of-the-art as systems that reliably identify all clinically significant interactions.