Biology Reference
In-Depth Information
four objective measures: two aggregate measures related to a simple urn model
of the AERS database, and two case-specific measures based on the numbers of
concomitant drugs listed in the database. These five association measures are de-
scribed in Secs. 15.2 through 15.4, and their drug-dependent interrelationships
are discussed briefly in Sec. 15.5.
The strength of this drug dependence motivated the clustering case study pre-
sented in Secs. 15.6 and 15.7, based on 36 drugs randomly selected from the
AERS database and 15 others of independent interest. The five association mea-
sures noted in the previous paragraph were then computed for each drug and the
100 most frequently occurring adverse events in the AERS database. The at-
tributes used for clustering were the correlations between the different association
measures, along with the mean value of the subjective association measure. Us-
ing a permutation-based approach to determine the number of clusters [16] and
a stepwise variable selection procedure, three clusters were identified: a “high-
blame” cluster, characterized by a large average subjective association value but
low correlations between this association measure and both of the objective ag-
gregate association measures; a “low-blame” cluster, characterized by a small av-
erage subjective association value, again with low correlations between subjective
and aggregate association measures; and an “appropriate blame” cluster, charac-
terized by a moderate average subjective association measure but relatively high
correlations between the subjective and aggregate measures.
The results presented here do not fully define the index of blame that motivated
the work, but they do suggest that the approach described here goes in the right
direction. For example, one reason for including the illegal drug cocaine in the
study was the expectation that it would generally have a high subjective blame,
representing a “positive control” for the classification; indeed, cocaine appears
consistently in the high-blame group in all of the clusterings described here. Also,
it is encouraging to note that all four of the statins considered (atorvastatin, lovas-
tatin, pravastatin, and simvastatin) consistently appear together in the appropriate
blame cluster. Similarly, it is not surprising that aspirin — a nonprescription drug
in wide use for a very long time — belongs to the low-blame group in all cluster-
ings. Indeed, it has recently been argued that, due to its long history of use, the
adverse event profile for aspirin has often been overlooked by medical praction-
ers, even though it is quite complicated and deserves careful clinical review [20].
More generally, it appears that the drugs in the low blame group identified here
are mostly drugs like aspirin that have been widely used for a variety of conditions
and have a long history of use.
At least four extensions of the results presented here suggest themselves. The
first would be to expand the case study to additional association measures.
In
Search WWH ::
Custom Search