Relating Subjective and Objective Pharmacovigilance Association Measures - Clustering Challenges in Biological Network - page 297

gastrointestinal haemorrhage nos

myocardial infarction

haemoglobin decreased
chest pain

death
maternal drugs affecting foetus
death nos

1 2 3 4

Reporting Ratio

Fig. 15.8. Adverse event plot for the drug aspirin, based on the 100 most frequently occurring ad-
verse events, with the point sizes proportional to the suspect fraction S ab .

It is instructive to compare the appearance of the adverse event plots obtained
for drugs from each of these groups with point sizes proportional to the subjective
association measure S ab . Fig. 15.8 presents this result for the drug aspirin from
cluster no. 1 (the “low-blame group”). This plot should be compared with Figs.
15.2 and 15.3 discussed earlier, for the drugs infliximab (from cluster no. 3, the
“high-blame group”)andfluoxetine (from cluster no. 2, the “appropriate blame
group”), respectively. It is clear from these plots that the points are uniformly
large for infliximab, indicating a large S ab value, mostly independent of U ab and
R ab , and generally small for aspirin, indicating a small S ab value, again mostly
independent of U ab and R ab . In contrast, the point sizes vary significantly for
fluoxetine, indicating that S ab varies significantly with U ab and R ab for this drug.
The informal verbal descriptions given above for each of the three partitions
in the basic clustering are represented graphically in Fig. 15.9, which gives side-
by-side boxplot summaries by cluster for each of three variables on which this
clustering is based. The left-most three boxplots show how the mean S ab value
varies for each cluster, demonstrating that this value is consistently larger for clus-
ter 3 than for either of the other two clusters. Also, while the total range of S ab
values is essentially identical for clusters 1 and 2, the middle 50% of these values