Safety Data Mining:
Background and Current Issues
Ramin Arani, PhD
Safety Data Mining
Global Biometric Science
Bristol

Myers Squibb Company
SAMSI: July, 2006
Outline
Rationale for
Pharmacovigilance
AERS Data Base
Data base issues
Methodologies
BCNN (WHO)
MGPS (FDA)
Summary
Challenges and Opportunities
Pharmacovigilance

Rationale
Information obtained prior to first marketing is inadequate to cover all
aspects of drug safety:
tests in animals are insufficiently predictive of human safety,
in clinical trials patients are selected and limited in number,
conditions of use in trials differ from those in clinical practice,
duration of trials is limited
information about rare but serious adverse reactions, chronic
toxicity, use in special groups or drug interactions is often not
available.
Pre Approval Data

Controlled

Limited # Pts

Safety data not mature
Post Approval Data

Real life ; uncontrolled

Off label use

Generic

Solicited Safety
Data

Unsolicited Safety
Data
Population
Subjects for
approval
Pharmacovigilance

Rationale
Spontaneous AE Reports
Safety information from clinical trials is incomplete
°
Few patients

rare events likely to be missed
°
Not necessarily ‘real world’
Need info from post

marketing surveillance & spontaneous reports
Pharmacovigilance by reg. agencies & mfrs carried out.
Long history of research on issue
°
Finney (MIMed1974, SM1982)
Royall (Bcs1971)
°
Inman (BMedBull1970)
Napke (CanPhJ1970)
Issues
Incomplete reports of
events
, not necessarily reactions
How to compute effect magnitude
Many events reported, many drugs reported
Bias & noise in system
Difficult to estimate
incidence
because no. of pats at risk, duration of
exposure seldom reliable
Appropriate use of computerized methods, e.g., supplementing
standard pharmacovigilance to identify possible signals sooner

early warning signal
Safety Signal:
Reported information on a possible causal relationship between
an adverse event and a drug.
Pharmacovigilance

Definition
Phamacovigilance
Set of methods that aim at identifying and quantitatively
assess the risks related to the use of drugs in the entire
population, or in specific population subgroups
Adverse Drug Reaction
A
response
to a drug which is harmful and unintended, and which
occurs at doses normally used.
AERS Database
Database Origin 1969
SRS until 11/1/97; changed to AERS
3.0 million reports in database
All SRS data migrated into AERS
Contains Drug and "Therapeutic" Biologic Reports
exception = vaccines (VAERS)
Source of AERS Reports
Health Professionals, Consumers / Patients
Voluntary : Direct to FDA and/or to Manufacturer
Manufacturers: Regulations for Postmarketing Reporting
AERS Limitations
Different populations, Co

morbidities, Co

prescribing, Off

label
use, Rare events
Report volume for a drug is affected by, volume of use,
publicity, type and severity of the event and other factors,
therefore the reporting rate is not a true measure of the rate or
the risk
An observed event may be due to the indication for therapy
rather than the therapy itself; therefore observed associations
should be viewed as signal, and causal conclusions drawn
with caution
Examples
Claritin and arrhythmias
(channeling and need for detailed
data not in data base)
Increased number of reports due to preexisting
condition. Selection of high risk patients for the drug
deemed safest for them.
Prozac and suicide
(confounding by indication) Large
increase in reports following publicity and stimulated
reporting
The Pharmacovigilance Process
Detect Signals
Traditional
Methods
Data
Mining
Generate Hypotheses
Refute/Verify
Type A
(Mechanism

based)
Type B
(Idiosyncratic)
Insight from
Outliers
Estimate
Incidence
Public Health
Impact, Benefit/Risk
Act
Inform
Change Label
Restrict use/
withdraw
Methodologies
Finding “Interestingly Large” Cell Counts
in a Massive Frequency Table
Rows and Columns May Have Thousands of Categories
Most Cells Are Empty, even though
N
++
Is very Large
Only 386K out of 1331K Cells Have
N
ij
> 0
174 Drug

Event Combinations Have
N
ij
> 1000
No. Reports
AE
1
…
AE
n
Total
Drug 1
N
11
…
N
1n
N
1+
:
:
N
ij
:
:
Drug m
N
m
1
…
N
mn
N
m+
Total
N
+1
…
N
+n
N
++
Method

Basics
Endpoint: No of AEs
Most use variations of 2

way table statistics
No. Reports
Target
AE
Other
AE
Total
Target Drug
a
b
a+b
Other Drug
c
d
c+d
Total
a+c
b+d
n
Some possibilities
Reporting Ratio:
E(a) = (a+b)
(a+c)/n
Proportional Reporting Ratio:
E(a) =
(a+b)
c / (c+d)
Odds Ratio:
E(a) = b
c / d
OR > PRR > RR when a > E(a)
Basic idea:
Flag when
R =
a
/E(a) is
“large”
Bayesian Approaches
Two current approaches: DuMouchel & WHO
Both use ratio n
ij
/ E
ij
where
n
ij
= no. of reports mentioning both drug i & event j
E
ij
= expected no. of reports of drug i & event j
Both report features of posterior dist’n of ‘information criterion’
IC
ij
= log
2
n
ij
/ E
ij
= PRR
ij
E
ij
usually computed assuming drug i & event j are mentioned
independently
Ratio > 1 (IC > 0)
combination mentioned more often than
expected if independent
WHO (Bate et al, EurJClPhrm1998)
‘Bayesian Confidence Neural Network’ (BCNN) Model:
n
ij
= no. reports mentioning both drug i & event j
n
i+
= no. reports mentioning drug i
n
+j
= no. reports mentioning event j
Usual Bayesian inferential setup:
Binomial likelihoods for n
ij
, n
i+
, n
+j
Beta priors for the rate parameters (r
ij
, p
i
, q
j
)
WHO, cont’d
Uses ‘delta method’ to approximate variance of
Q
ij
= ln r
ij
/ p
i
q
j
= ln 2
IC
ij
However, can calculate exact mean and variance of Q
ij
WHO measure of importance = E(ICij)

2 SD(ICij)
Test of signal detection predictive value by analysis of signals 1993

2000: Drug Safety 2000; 23:533

542
84% Negative Pred Val, 44% Positive Pred Val
Good filtering strategy for clinical assessment
WHO, cont’d
WHO. (Orre et al 2000)
IC
D
P
A
P
D
A
P
D
A
P
D
A
I
,
log
,
,
1
,
0
,
log
log
,
log
log
,
,
,
1
1
1
2
1
i
k
d
k
i
k
i
i
k
d
k
i
k
i
i
d
k
i
k
i
d
i
i
i
i
i
if
A
P
d
P
A
d
P
A
P
A
P
d
P
A
d
P
A
P
A
P
d
P
A
d
P
A
P
d
P
A
d
P
A
P
d
P
A
d
P
A
P
d
d
A
P
D
A
P
k
i
k
i
k
i
i
Let
A
denote adverse events and
D
denote the drug.
Mutual information
I
(A,D) is a measure of association
WHO, cont’d
DuMouchel (AmStat1999)
E
ij
known, computed using stratification of database

n
i+
(k)
= no. reports of drug i in stratum k
n
+j
(k)
= no. reports of event j in stratum k
N
(k)
= total reports in stratum k
E
ij
=
k
n
i+
(k)
n
+j
(k)
/ N
(k)
(E (n
ij
) under independence)
n
ij
~ Poisson(
ij
)

interested in
ij
=
ij
/E
ij
Prior dist’n for
= mixture of gamma dist’ns:
f(
; a
1
, b
1
, a
2
, b
2
,
) =
g(
; a
1
, b
1
) + (1
–
) g(
; a
2
, b
2
)
where
g(
; a, b) = b (b
)
a
–
1
e

b
/
(a)
DuMouchel, cont’d
Estimate
,
a
1
, b
1
, a
2
, b
2
using Empirical Bayes

marginal dist’n of
n
ij
is mixture of negative binomials
Posterior density of
ij
also is mixture of gammas
ln
2
ij
= IC
ij
Easy to get 5% lower bound (i.e. E(IC
ij
)

2 SD(IC
ij
) )
The control group and the issue
of ‘compared to what?’
Signal strategies, compare
a drug with itself from prior time periods
with other drugs and events
with external data sources of relative drug usage and
exposure
Total frequency count for a drug is used as a relative surrogate for
external denominator of exposure; for ease of use, quick and
efficient;
Analogy to case

control design where cases are specific AE term,
controls are other terms, and outcomes are presence or absence of
exposure to a specific drug.
Other useful metrics and methods
Chi

square statistics
P

value type metric

overly influenced by sample size
Modeling association through directly Multivariate Poisson dist
Incorporation of a prior distribution on some drugs and/or
events for which previous information is available

e.g. Liver
events or pre

market signals
Interpreting the Signal Through
the Role of Visual Graphics
Four examples of spatial maps that reduce the scores to
patterns and user friendly graphs and help to interpret
many signals collectively
Example 1
A spatial map showing the “signal scores” for the most
frequently reported events (rows) and drugs (columns) in
the database by the intensity of the empirical Bayes signal
score (blue color is a stronger signal than purple)
Example 2
Spatial map showing ‘fingerprints’ of signal scores allowing one
to visually compare the complexity of patterns for different drugs
and events and to identify positive or negative co

occurrences
Example 3
Cumulative scores and numbers of reports according to the
year when the signal was first detected for selected drugs
Example 4
Differences in paired male

female signal scores for a specific
adverse event across drugs with events reported (red means
females greater, green means males greater)
Summary
1.
There is NO Golden Standard method for signal detection.
2.
The signals become more stable over time, however there is a
limited time window of opportunity for signal detection.
3.
Use Time

slice evolution of signal.

Fluctuation might reveal external risk factors.

Robustness can be assessed.
4.
Consider other endpoint such as time to onset, duration of
event, etc.
5.
For spontaneous case reports, the means to improve content is
to standardize and improve intake
6.
Data mining likely will generate many false positives and
affirmations of what was previously known
7.
Causality assessments should largely be reserved refining
important signals
Challenges in the future
More real time data analysis
More interactivity ( Visual Data mining, e.g. ggobi )
Linkage with other data bases to control the bias
inherent in data base
Quality control strategies (e.g. Identifying duplicates
Methods to reduce the false positive and negative?
Comments 0
Log in to post a comment