associations and do not necessarily imply a clinical
or a causal association.
As discussed, there are several disproportional-
ity analyses that can be conducted, but essentially
an observed rate is compared to an expected rate.
Following the calculation of an EBGM or a PRR,
there is an ability to generate a case series and then
to characterize that case series. As a result of the
surveillance process, hypotheses are generated.
These hypotheses may then need to be evaluated
using additional quantitative methods as appropri-
ate, looking at the company’s database, or requir-
ing stimulated reporting, or enhanced surveillance,
or epidemiological studies to try and evaluate these
hypotheses.
Thus, it is clear that the safety surveillance pro-
cess is an iterative one. It looks at multiple data
sources, whether screening large regulatory data-
bases, looking at company databases or looking at
manufacturing Lot related AEs for patential pro-
blems. The surveillance process screens the data
using both the intraproduct and the interproduct
methods. The object is to identify topics for further
review to develop case definition, to compile a case
series and then to characterize that case series.
40.4 Data mining in safety
surveillanceData mining is used in the review of safety surveil-
lance data to try and detect strong, consistent asso-
ciations that occur at higher-than-expected
frequencies. Data mining usually uses AE safety
databases that lack denominator data. It detects
frequency of drug event combinations in post-
marketing reports. It also determines the relative
frequency that the drug–event combinations are
reported for drug X than for any other drug. Data
mining attempts to quantify the strength of poten-
tial drug–event association, whereby signal scores
are calculated and represent the relative reporting
rate for AEs.
Data mining does not equal data dredging. It is a
systematic screening for drug–event combinations
that are being reported disproportionately. It is
essentially a quantitative signal detection method.
The data mining method that is currently being
used widely in the United States, by both the
FDA and the pharmaceutical industry, is the
MGPS, which adjusts for the multiplicity of
drugs and events per record. The MGPS generates
an EBGM, which is an estimate of the relative
reporting ratio. It is the ratio of the observed over
the expected counts. A 90% confidence interval is
calculated around the EBGM covering the lower
5% and the upper 95% of the confidence interval.
Amongst the major challenges that data mining
has had is the belief that this is simply data dred-
ging and that this is not a worthwhile scientific
endeavor. However, data mining is a scientific,
statistically valid method that encompasses a
quantitative computer-assisted method of trying
to determine safety signals. Both of the FDA’s
post-marketing safety surveillance databases, the
Adverse Events Reporting System (AERS) and
the Vaccines Adverse Events Reporting System
(VAERS) are used in safety signaling and data
mining, along with the World Health Organiza-
tion’s (WHO) AE database. The FDA’s databases
contain all US reports along with serious, unla-
beled reports from outside the United States. The
WHO database contains reports from more than
65 national authorities, including the FDA’s data-
bases.
The FDA’s Spontaneous Reporting System(SRS)
was in operation from 1968 to October 1997. The
reports were transitioned to the AERS, which has
been used from October 1997 to the present. A
publicly released version can be purchased on a
quarterly basis. It is a passive surveillance system
where direct volunteer reporting accounts for 10%
of reports from healthcare professionals and consu-
mers. Ninety percent of reports in the FDA’s post-
marketing safety surveillance databases come from
pharmaceutical companies, as they are mandated by
regulations to report AEs that they receive. The
combined SRSþAERS database currently con-
tains more than 2.7 million reports and is growing
rapidly, at about 465, 000 reports on anannual basis.
The number of reports has more than doubled in the
last 10 years and because the FDA is interested in
serious, unlabeled reports, that has grown as a per-
centage of the total number of reports submitted to
the FDA.548 CH40 DATA MINING