Relating Subjective and Objective Pharmacovigilance Association Measures - Clustering Challenges in Biological Network - page 280

medical conditions is (or is not) suspicious enough to report. Nevertheless, ob-
jective assessments of association are possible by comparing the relative frequen-
cies with which different drug/adverse event combinations occur in the database.
These aggregate assessments are based on the statistical measures discussed in
Sec. 15.2. In addition, the AERS database also provides subjective association
data, described in Sec. 15.3, and the numbers of drugs listed in each ISR provide
the basis for defining two case-specific objective association measures, discussed
in Sec. 15.4. It is demonstrated with two simple examples in Sec. 15.5 that the
relationships between these five association measures is strongly drug-dependent.
The question of ultimate interest is how to define an index of blame that de-
scribes the tendency for a drug to be blamed for adverse reactions to a significantly
greater or lesser extent than is warranted on the basis of the objective association
measures. As a first step toward this goal, this paper presents the results of a clus-
ter analysis of drugs, using correlations between different association measures
over a fixed set of adverse events as their primary attributes. Detailed descriptions
of the clustering method used, the drugs and adverse events considered, and the
results obtained are presented in Sec. 15.6, and an interpretation of these results
is given in Sec. 15.7. It is seen that the drugs considered here cluster naturally
into three groups: a “high-blame” group, where subjective association measures
between the drug and most adverse events are high, largely independent of the
objective evidence for this association; a “low-blame” group, where subjective as-
sociation measures are low, again largely independent of objective evidence; and
an “appropriate blame” group, where subjective and objective association mea-
sures are in reasonable agreement. This result is discussed further in the summary,
given in Sec. 15.8.
15.2. Aggregate Associations
Many different algorithms have been proposed for pharmacovigilance analysis.
One of the earliest was the use of the Proportional Reporting Ratio (defined in
Eq. (15.3) below) to detect unusually strong drug/adverse event associations ad-
vocated by Finney in 1974 [5]. Refinements of this approach were proposed
by Evans et al. [4] to address some of the limitations described below, recog-
nition of which also led to the development of Bayesian alternatives like the
Gamma Poisson Shrinker (GPS) algorithm and its extensions by DuMouchel and
co-workers [2, 3, 6] and the Bayesian Confidence Propagation Neural Network
(BCPNN) method described by van Puijenbroek et al. [17]. In addition, non-
statistical alternatives have been developed, like the optimization-based method
that Mammadov and co-workers have applied to the Australian Adverse Drug