Transcript WebVDME/MGPS

Detecting Multi-Item Associations and
Temporal Trends Using the
WebVDME/MGPS Application
DIMACS Tutorial on Statistical
and Other Analytic Health
Surveillance Methods
18 June 2003
Richard Ferris
Pharmaceutical post-marketing surveillance

Companies and regulatory agencies collect databases of
spontaneous adverse reaction reports

Relevant exposure data not readily available (the “denominator
problem”)

Can drug-event combinations of potential interest be identified
from internal evidence alone?

Approach:
–
–
Use an internally defined “denominator”
Construct set of “expected” counts using a stratified independence model
Computation of Expected Counts

The expected count for a given drug-event combination is
determined by the overall count for the particular drug (across
all events) and the overall count of the particular event (across
all drugs)

For example, if 2% of all reports have PROZAC as a drug, and
3% of all reports have RASH as an event, then one would
expect that 0.06% (0.02*0.03) of the reports will include this
combination (PROZAC in combination with RASH)

(MGPS carries out this computation separately for each distinct
“stratum” and sums the strata-specific expected counts to
obtain an overall expected count)
1.
Comparing Observed and Expected Counts:
Relative Reporting Rate

Relative Report Rate (RR):
RRij = Nij / Eij

Easy to interpret, easy to compute

Statistically unstable if N is small or E is very small

The following all have RR = 100:
–
–
–
–
N = 1000,
N = 100,
N = 10,
N=
1,
E = 10
E= 1
E = 0.1
E = 0.01
2.
Comparing Observed and Expected Counts:
Statistical Significance

What is the probability that Nij would be observed by chance
(“sampling error”) when expected value is Eij ? (p-value for
testing a null hypothesis)

Harder to interpret (not expressed in same units as RR)

Results in computation of absurdly small probabilities that have
no meaning
–

N=100, E=1 produces 10-158 !
Small RR can be very significant (small p-value) when sample
size is very large:
–
–
N = 2000, E = 1000, RR = 2
N = 10, E = 0.1, RR =100
is more “significant” than
3.
Comparing Observed and Expected Counts:
Empirical Bayes Multi-Item Gamma Poisson Shrinker

Try for best of both previous approaches
–
–

Focus on the distribution across the set of drug-event combinations of
the ratios:
–





interpretability of relative rate
adjust properly for sampling variation
Estimate lij = mij /Eij , where Nij ~ Poisson(mij )
Fit a parameterized “prior distribution” function (mixture of two gamma
functions) to the empirical distribution of the l’s
Find posterior distribution of l after observing N = some value n
Use this to obtain posterior estimate of expectation value of l given
observation of Nij
This posterior estimate is what we call EBGM (Empirical Bayes
Geometric Mean); also get lower and upper 95% confidence bounds
(EB05, EB95).
EBGM is termed the “shrinkage estimate” for RR
Multi-Item Associations
vs. Pairwise Associations

Consider the case of an item triplet; e.g. 2 drugs and an event

RRijk = Nijk/Eijk where Eijk is based on independence model

EBGMijk = shrinkage estimate of RRijk

Suppose a particular itemset (drug A, drug B, event C = kidney failure) is
unusually frequent (EBGM for the triplet is >> 2)

Important to ask:
–
–

OR
Compare Empirical Bayes estimate of the frequency count of the triplet to the
prediction from the all-2-factor log-linear model
–
–
–
–

Is this merely the result of one or more of the pairs (AB, AC, BC) being unusually frequent?
Is this a drug-drug interaction
EXCESS2 = (EBGM * E ) – EAll2F
E is the expected count from independence
Computation of EAll2F uses shrinkage estimates of pairwise counts
EXCESS2 is an estimate of how many “extra” cases were observed over what was expected using
the all-2-factor model
Alternate approach: Define Eijk from predictions of all-2-factor model in which
case resulting EBGM directly measures divergence of observed count from all2-factor prediction
Health Authority Adoption of Signal
Detection Technologies

FDA
–
CDER:




–
CBER:

–

initial GPS implementation (VAERS)
CRADA between Lincoln and FDA to further develop methodology and
tools
CDC
–
–

Experimented in Office of Biostatistics with GPS for several years
Validated GPS
Moving to production
Have published data mining results on internal web for almost all products
Collaborative GPS methodology development with FDA
Includes simulation capability
WHO Uppsala Monitoring Centre
–
Production safety signal generation mechanism using BCPNN
FDA/GPS Validation Activities

Positive controls
–

Negative controls
–

Examine data mining results for drug-event combinations corresponding to
known “labeled” adverse reactions
Examine data mining results for several drugs (with differing safety profiles)
given for the same indication
“Roll back” database in time to determine when method would
have provided first signal
Databases of Spontaneous AE Reports

FDA Spontaneous Report System (SRS)
–
–
–
–

FDA Adverse Event Reporting System (AERS)
–
–
–
–
–
–
–
–

Post-Marketing Surveillance of all Drugs since 1969
Dates from mid-60’s thru 1997
1.5 Million Reports
Encoded in COSTART
US cases, serious unlabeled events from all manufacturers.
All products sold in the US ~5000 Rx’s
Replaced SRS in 1997
Reactions coded as MedDRA PTs
Quarterly Updates, 4-6 month delay
Drugs are Verbatim
Includes initial and some follow-up reports
Includes Demographics, Reactions, Drugs, Outcomes, etc.
FDA/CDC Vaccine Adverse Events (VAERS)
–
Stricter Laws for Vaccine Adverse Event Reporting
Signal Detection Demonstration
Using VAERS Data
“Significant” EBGM and even
extremely conservative EB05
with small N
Simple Rankings
by Signal Strength
Evolution of Signals
Over Time
Multi-Symptom Syndromes
(Higher Order Associations)
The “Serotonin Syndrome”

Could MGPS be used to identify unknown syndromes?

Try mining the AERS data for “significant” event triples using a
known syndrome.

"The symptoms of the serotonin syndrome are: euphoria,
drowsiness, sustained rapid eye movement, overreaction of the
reflexes, rapid muscle contraction and relaxation in the ankle
causing abnormal movements of the foot, clumsiness,
restlessness, feeling drunk and dizzy, muscle contraction and
relaxation in the jaw, sweating, intoxication, muscle twitching,
rigidity, high body temperature, mental status changes were
frequent (including confusion and hypomania - a "happy drunk"
state), shivering, diarrhea, loss of consciousness and death.
(The Serotonin Syndrome, AM J PSYCHIATRY, June 1991)
Using Simulation to Test
the Signal Detection Process
Interpreting Simulation Parameters
Outcome
Yes
No
Yes
R
No
Q-R
Q
Exposure
P-R
1-P-Q+R
1-Q
P
1-P
1
1.
As R  P and (Q-R)  (1-P) => “No Signal”
2.
As R  P and (Q-R) << (1-P) => “Strong Signal”
3.
When R << P and (Q-R)(1-P) => “No Signal”
4.
When R << P and (Q-R) << (1-P) => “Rare event”
Using Simulation to Create a Receiver Operating
Characteristic (ROC) Curve for EBGM

An ROC curve displays the true-positive rate
(sensitivity) versus the false-positive rate
(1 – specificity) for a statistic

Ran a 20 iteration simulation using P = 0.003
Q = 0.001 and R = 0.00003 (RR = 10) to check the
true-positive rate

Ran a 20 iteration simulation using P = 0.003,
Q = 0.001 and R = 0.0003 (RR = 1) to check the
false-positive rate
ROC Curve Based on Simulated Injection of Signals
Simulating a Rare Event

Sample 100,000 records from VAERS data

Set P = 0.003, Q = 0.001, R = 0.00003

Iterate 20 Monte Carlo simulations

Expect (on average):
–
–
–
–
–
0.003 x 100,000 = 300 “Rare Exposures”
0.001 x 100,000 = 100 “Rare Outcomes”
0.00003 x 100,000 = 3 “Rare Exposure + Rare Outcome”
combinations
E = (300 x 100) / 100,000 = 0.3
RR = 3/ 0.3 = 10
Base Simulation on VAERS Data
Sample Cases From VAERS
Sample 100,000 Cases
P = 0.003
Q = 0.001
R = 0.00003
20 Monte Carlo Iterations
RareExposure
Expected N = 300
RareOutcome
Expected N = 100
RareExposure + RareOutcome
Expected N = 3
Expected RR = 10
Technical Details

William DuMouchel. Bayesian Data Mining in Large Frequency
Tables (with Discussion). The American Statistician (1999) pp
177-190.

William Dumouchel and Daryl Pregibon. Empirical Bayes
Screening for Multi-Item Associations. Proceedings of KDD
2001.
Methodology History and Key
Contributors

Stephan Evans
–
–
–

Bate, Lindquist, Edwards et. al.
–
–

MCA, UK
Proportional reporting ratio (PRR) with Chi 2 analyses
Simple, highly intuitive, can be calculated by hand
WHO Uppsala Monitoring Centre
Bayesian neural network method for adverse drug reaction signal generation
Ana Szarfman, FDA (CDER) and Bill DuMouchel (ATT)
–
Empiric Bayes, more robust than PRR for small n



–
Multidimensional analyses possible

–
MGPS method: statistical parameter is EGBM
William DuMouchel. Bayesian Data Mining in Large Frequency Tables (with Discussion). The
American Statistician (1999) pp 177-190.
William Dumouchel and Daryl Pregibon. Empirical Bayes Screening for Multi-Item
Associations. Proceedings of KDD 2001.
Interactions, gender and other demographic associates, syndrome identification
Can directly compare EBGM values of different drugs, as well as for a specific drug
Key Contributors (continued)

WHO Collaborating Center for Internat’l Drug Monitoring: M
Lindquist, M Stahl, A. Bate, R. Edwards, RH Meyboom.
–
–

L. Gould . Comparison and refinement of Bayesian approaches
for evaluating spontaneous reports of ADRs. DIA Annual
meeting, July 2001, (Denver)
–

Bayesian confidence propagation neural network (BCPNN) . Information
Component (IC) statistic is the measure of the strength of D:E relationship
Iterative approach
EB vs BCPNN = similar results
Thakrar, BT, Blesch, KS, Sacks, ST, Wilcock, K (2001)
–
–
(ISPE, Pharmacoepid. & Drug Safety 10),
PRR vs. EB= similar sensitivity, EB better at ranking events based on small
N.