Relating Subjective and Objective Pharmacovigilance Association Measures - Clustering Challenges in Biological Network - page 302

particular, one of the features of the AERS database that is not addressed in the
present study is the fact that while each drug listed for an ISR has a unique sub-
jective classification as Primary Suspect, Secondary Suspect, Interacting, or Con-
comitant, this designation does not indicate whether the classification refers to all
of the adverse events listed for that ISR, or only a subset of these adverse events
(and, if so, which subset). This point is important since, as noted, a typical ISR
lists multiple drugs and multiple adverse events. Thus, one area to explore is
the possibility of defining additional association measures based in part on the
number of adverse events listed for an ISR, analogous to the mean number of
concomitant medications µ ab or the pure play fraction ψ ab defined in Sec. 15.4
for drugs. The objective would be to explore the influence of the number of ad-
verse events listed for an ISR on clustering results like those presented here. The
second useful extension of the results presented in this paper would be to expand
the case study to more drugs and possibly more adverse events. This expansion
would begin to address the question of whether the clustering of drugs into “low
blame,”“appropriate blame,” and “high blame” classes is adequate, or whether
additional classifications become necessary as more drugs are added, possibly in-
volving more variables. In particular, it would be interesting to see whether the
four cluster solution discussed in Sec. 15.7 based on five variables increases in
statistical significance on the addition of more drugs. A third possible exten-
sion would be to consider the drugs assigned to each of the three groups in the
basic clustering developed here, comparing them on the basis of other possible
explanatory variables like their frequency of occurrence in the AERS database,
or the time since their introduction into the marketplace. The objective of this
extension would be to determine whether it is possible to reliably assign drugs to
different “blame clusters” on the basis of these other variables without perform-
ing the cluster analysis presented here. In particular, if it were possible to classify
drugs apriori into these groups, this classification could be used as a basis for
requiring additional evidence in classifying a “high blame drug” as suspect, or
classifying a “low blame drug” as concomitant. Finally, a fourth extension of
the work presented here would be to carefully examine anomalous cases like the
drug ciprofloxacin: the three other drugs from the same class considered in this
case study (gatifloxacin, levofloxacin, and moxifloxacin) are all assigned to the
high-blame cluster, while ciprofloxacin is assigned to the low-blame cluster. One
possible explanation for this difference is that ciprofloxacin is a much older drug
than these other three, but this case needs to be examined further before any defini-
tive conclusions can be drawn. Ultimately, it is hoped the work presented here and
extensions like those just described will lead to an evidence-based approach to as-