ArticlePDF Available

Abstract

Objective: Signal detection is a technique in pharmacovigilance for the early detection of new, rare reactions (desired or undesired) of a drug. This study aims to compare and appraise the performance of data mining algorithms used in signal detection. Method: Most commonly used three data mining algorithms (DMAs) (Reporting Odds Ratio (ROR), Proportional Reporting Ratio (PRR) and Information Component (IC)) were selected and applied retrospectively in USFDA Adverse Event Reporting System database to detect five confirmed Drug Event Combinations. They were selected in such a way that the drug is withdrawn from the market or label change between 2006-2015. A value of ROR-1.96SE>1, PRR≥2, χ2>4 or IC- 2SD>0 were considered as the positive signal. The data mining algorithms were compared for their sensitivity and early detection. Result: Among the three data mining algorithms, Information Component was found to have a maximum sensitivity (100%) followed by Reporting Odds Ratio (60%) and Proportional Reporting Ratio (40%). Sensitivity associated with the number of reports per drug event combination and early signal detection suggested that information component needs comparatively fewer reports to show positive signal than the other two data mining algorithms. ROR and PRR showed comparable results. Conclusion: Early detection of a reaction is possible using signal detection technique. Information component was found to be sensitive method compared with other two data mining algorithms in FDA Adverse Event Reporting System database. As the number of reports of drug event combination increased, the sensitivity and comparability of data mining algorithm also increased.
J Young Pharm, 2018; 10(4): 444-449
A multifaceted peer reviewed journal in the eld of Pharmacy
www.jyoungpharm.org | www.phcog.net
Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018 444
Original Article
INTRODUCTION
e US Food and Drug Administration (FDA) has maintained the
Adverse Events Reporting System (AERS) database since 1968 with the
intention of continuous monitoring of drugs during the post-marketing
surveillance.1 AERS database is a collection of suspected Adverse Event
(AEs) reports from pharmaceutical companies, consumers and health-
care professionals.2 To date, FDA AERS database contains more than
10 million AE reports and receive nearly half a million each year. Many
approaches have been adopted into post-marketing studies, including
prescription-event monitoring (PEM), spontaneous reporting, case
review, observational study and drug utilization review.
Compared to clinical trials and traditional epidemiologic studies, the
computer-assisted Data Mining Algorithms (DMAs) are relatively new
and characterized by providing a fast and ecient way of detecting
possible Adverse Drug Reactions (ADRs) signal.3 Several DMAs have
been well described in the literature, mainly including the Reporting
Odds Ratio (ROR),4 Multi-item Gamma Poisson Shrinker (MGPS),5 the
Proportional Reporting Ratio (PRR),6 and the Information Component
(IC).7 By integrating such computer-aided techniques, incorporating
statistical analyses and centralizing dierent data resources, it is not only
possible to minimize the human eorts and errors but also assist the
regulatory bodies and safety councils.8
A.M. Wilson9 dened ‘data mining’ as ‘the use of statistical techniques,
such as disproportionality measures for database or large information
sources for extracting an unknown information’. At present, three major
DMAs such as the PRR of the Netherlands, the ROR of the United Kingdom
and the IC of the WHO are widely used.9-10
Speaking of the DMAs, one of the frequently discussed and yet to be
resolved question is which algorithm has superior performance. e
absence of gold standard,11 enormous ADR reports, dierent coding
systems, a wide range of data mining processes and structural dissimi-
larities of databases made the comparisons across the DMAs dicult.
Limited studies have been conducted to compare the DMAs thus far.
is study aims to compare and appraise the performance of signal
detection techniques used in data mining.
METHODOLOGY
Most commonly used three DMAs (ROR, PRR and IC) were selected
based on a literature survey. DMAs were applied retrospectively (Table 1
and 2) in US FDA AERS database to detect ve conrmed Drug Event
Combinations (DEC) (Table 3). e DEC was selected in regard to the
withdrawal of the drug from the market between 2006-2015 or the
change in labelling criteria or black box warning of the drug during the
time period of 2006-2015. e time period is important because the data
available in USFDA AERS database for signal detection is from 2006.
Data source
AE reports from the FDA AERS database were used for the study. It is a
surveillance program used for detecting serious AEs that have not been
identied during premarketing analysis.12
A Comparative Study of Data Mining Algorithms used for Signal
Detection in FDA AERS Database
Viswam Subeesh1,*, Eswaran Maheswari2, Ganesan Rajalekshmi Saraswathy2, Ann Mary Swaroop3, Satya Sai Minnikanti3
1Research Scholar, Department of Pharmacy Practice, Faculty of Pharmacy, M.S Ramaiah University of Applied Sciences, Bengaluru, Karnataka, INDIA.
2Professor, Department of Pharmacy Practice, Faculty of Pharmacy, M.S Ramaiah University of Applied Sciences, Bengaluru, Karnataka, INDIA.
3Associate Professor, Department of Pharmacy Practice, Faculty of Pharmacy, M.S Ramaiah University of Applied Sciences, Bengaluru, Karnataka, INDIA.
ABSTRACT
Objective: Signal detection is a technique in pharmacovigilance for the
early detection of new, rare reactions (desired or undesired) of a drug. This
study aims to compare and appraise the performance of data mining
algorithms used in signal detection. Method: Most commonly used three
data mining algorithms (DMAs) (Reporting Odds Ratio (ROR), Proportional
Reporting Ratio (PRR) and Information Component (IC)) were selected and
applied retrospectively in USFDA Adverse Event Reporting System database
to detect ve conrmed Drug Event Combinations. They were selected in
such a way that the drug is withdrawn from the market or label change
between 2006-2015. A value of ROR-1.96SE>1, PRR≥2, χ2>4 or IC- 2SD>0
were considered as the positive signal. The data mining algorithms were
compared for their sensitivity and early detection. Result: Among the three
data mining algorithms, Information Component was found to have a
maximum sensitivity (100%) followed by Reporting Odds Ratio (60%) and
Proportional Reporting Ratio (40%). Sensitivity associated with the number
of reports per drug event combination and early signal detection suggested
that information component needs comparatively fewer reports to show
positive signal than the other two data mining algorithms. ROR and PRR
showed comparable results. Conclusion: Early detection of a reaction is
possible using signal detection technique. Information component was
found to be sensitive method compared with other two data mining
algorithms in FDA Adverse Event Reporting System database. As the number
of reports of drug event combination increased, the sensitivity and compa-
rability of data mining algorithm also increased.
Key words: Signal Detection, Data mining algorithms, FDA AERS Database,
Disproportionality Analysis, Pharmacovigilance.
Correspondence
Mr. Subeesh K Viswam, Research Scholar, Department of Pharmacy Practice,
Faculty of Pharmacy, M.S Ramaiah University of Applied Sciences, Bengaluru,
Karnataka, INDIA.
Phone: +91-9895843219
Email: subeeshkviswam@gmail.com
DOI: 10.5530/jyp.2018.10.97
This is an open access article distributed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License, which allows others
to remix, tweak, and build upon the work non-commercially, as long as the author is credited and the new creations are licensed under the identical terms.
Subeesh, et al.: A Comparative Study of Data Mining Algorithms
Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018 445
RESULTS
ROR, PRR and IC were applied on USFDA AERS database to detect ve
conrmed DEC and to compare the DMAs.
Propoxyphene
Propoxyphene was approved in 1957 and was withdrawn
from the market in 2010 owing to serious cardiac toxicity. A total of
366 DEC were reported from 2005Q1 to 2010Q4 in USFDA AERS
database (Table 4). ROR, 1.43 (95% CI, 1.26-1.52) and PRR, 2.8 (95% CI,
2.24- 3.32) showed positive signal from 2009Q1 and IC, 0.62 (95% CI,
0.06 -1.02) showed positive signal from 2007Q1 for the selected DEC.
PTs used for data mining were “cardiac arrest” and “cardio-respiratory
arrest”. Positive signals were highlighted with bold letters.
Sibutramine
Sibutramine was approved in 1997 as an oral anorexiant. Later in
2010, it was withdrawn from the market due to severe cardio and
cerebrovascular accidents and associated death. A total of 4182 DEC
were reported from 2005Q1 to 2010Q4 in USFDA AERS database
(Table 4). Only IC, 0.28 (95% CI, 0.05 – 0.54) showed a positive signal
for sibutramine from 2009Q1 whereas ROR and PRR failed to show a
positive signal for the given DEC in the specied duration. PT used for
data mining were “cerebrovascular disease” and “stroke”. Positive signals
were highlighted with bold letters.
Rosiglitazone
Rosiglitazone was approved in 1999 for diabetes mellitus and was with-
drawn from the market because of cardiac toxicity and associated death
in 2010. A total of 11839 DEC were reported from 2005Q1 to 2010Q4
in USFDA AERS database (Table 4). ROR, 1.4 (95% CI, 1-1.8) showed
positive signal from 2009Q3, PRR, 2.7 (95% CI, 2.1-3.6) from 2009Q4
and IC, 0.24 (95% CI, 0.02 - 0.5) showed positive signal from 2008Q3
for the given DEC in the specied period. PT used for data mining were
“Myocardial Infarction” (MI). Positive signals were highlighted with
bold letters.
Sitagliptin
Sitagliptin was approved in 2006 for diabetes mellitus. Recently (2012),
the labelling criteria of sitagliptin had changed to include a black box
warning of renal failure associated with sitagliptin. A total of 854 DEC
were reported from 2005Q1 to 2015Q2 in USFDA AERS database (Table
5). Only IC, 0.1 (95% CI, 0 – 0.23) showed positive signal (from 2011Q2)
whereas ROR and PRR failed to show positive signal. PT used for data
mining were “renal failure acute”, “creatinine renal clearance decreased”
and “blood creatinine increases”. Positive signals were highlighted with
bold letters.
Canagliozin
Canagliozin is approved in 2013 for the treatment of Diabetes Mel-
litus and in 2015 it was subjected to a change in labelling criteria to
incorporate safety warning of Urinary Tract Infection (UTI). A to-
tal of 566 DEC were reported from 2013Q2 to 2015Q4 in USFDA
AERS database (Table 5). IC, 0.78 (95% CI, 0.53- 0.98) and ROR, 1.42
(95% CI, 1.11-1.94) showed a positive signal at 2015Q3 and 2015Q4
respectively. PT used for data mining were “urinary tract infection.
Positive signals were highlighted with bold letters.
Sensitivity of DMAs
e sensitivity of three DMAs were assessed (Table 6) in accordance with
its potential to identify the signals prior to the withdrawal date. Out of
Table 1: The 2 × 2 table for the calculation of the signal.
Drug of Interest Other Drugs
ADR of interest A B
Other ADR C D
A: e number of reports containing both suspected drug and suspected ADRs
B: e number of reports containing drug of choice but with other ADRs
C: e number of reports containing the event of interest but with other medications
D: e number of reports concerning other medications and other ADRs
Table 2: Formula for the computation of signal.
Serial No. Measures Computation Threshold
1ROR ROR= (A/B)/(C/D)
SE
ABCD
.=+++
11
11
ROR-1.96SE>1
2PRR PRR=(A(A+C))⁄(B(B+D))
SE AACB
BD
=−
++++
11
11
PRR≥2
χ2>4
≥ 3 cases reported
3IC IC
pxy
pxpy
=log
(,)
()
()
2IC- 2SD>0
p(x) = Probability of a suspected drug being reported in a case report; p(y) =
Probability of a suspected event being reported in a case report; p(x, y) = Probability
that suspected drug and event simultaneously being reported in a case report
Table 3: Drug Event Combinations.
Sl. No. Drug Event Present status
1 Propoxyphene Serious toxicity to the
heart
Withdraw from
market
2 Sibutramine Cardio and
cerebrovascular
outcomes and death
Withdraw from
market
3 Rosiglitazone Serious cardiac toxicity Withdraw from
market
4Sitagliptin Renal failure Change in label
5 Canaglifozin Urinary Tract Infection Change in label
Source: FDA annual report
Study Procedure
FDA AERS database downloaded from USFDA ocial website. e
database converted into excel from text format for the ease of anal-
ysis and computation. Primary suspects and secondary suspect
case IDs of the drug of interest were noted from DRUG le. AE pertaining
to those case IDs were ltered and listed. Dierent parameters of the
DMA equation (Table 1) were computed and nally applied to the equation
(Table 2).
Statistical analysis
e threshold was predened as PRR of ≥2.0 with a Chi-squared test
of ≥4.0, at least three reports (n ≥3) of that preferred term(PT), IC with
IC-2SD>0 and ROR with ROR-1.96SE>1. e condence interval (95%)
were considered to be statistically signicant.13
Subeesh, et al.: A Comparative Study of Data Mining Algorithms
446 Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018
tively fewer reports to show positive signal than ROR and PRR. Parallel
to the above result, ROR and PRR are almost identical in their sensitivity.
DISCUSSION
In this study, the most commonly used three DMAs are ROR, PRR and
IC which are applied retrospectively in USFDA AERS database to detect
ve known and conrmed ADRs associated with the drug withdrawal
or change in labelling criteria. Further comparisons across the selected
DMAs were executed to identify the sensitivity by means of early detection
and number of reports.
Hitherto, there is no clearly dened method to compare the DMAs with
respect to their sensitivity or performance. e major drawback in
comparing the DMAs is the lack of golden standards.14 In E.P. van
Puijenbroek, A. Bate11 study, they compared the DMAs like PRR, Yule’s
Q and Chi-square with IC, which was considered as a golden standard by
them. Another study conducted by K. Kubota15 in Japanese spontaneous
reports, compared ve DMAs. According to Kubota et al., the number of
ve, IC showed maximum sensitivity (100%) followed by ROR (60%)
and PRR (40%).
The sensitivity of DMA based on Early Detection
DMAs were assessed for their ability for early detection of a signal
(Table 7). Index Date of Withdrawn/Label change (IDW/L) were identied
from literature or ocial websites of USFDA. e quadrant from which
DEC started showing positive signals were allocated as Index Date of
Detection (IDD). Early detection was quantied by IDD subtracted from
IDW/L. IC was found to be the most sensitive, as it detects positive signal
well before other DMAs and there is no remarkable dierence in sensi-
tivity of early detection between ROR and PRR.
Sensitivity associated with the number of reports per DEC
DMAs were assessed for their sensitivity with respect to the number of
reports required to show positive signal (Figure 1). Sensitivity associated
with the number of reports per DEC suggested that IC need compara-
Table 4: DMA of propoxyphene and reporting of cardiac arrest, Sibutramine and reporting of cerebrovascular disease and Rosiglitazone and reporting
of MI.
Propoxyphene and reporting of cardiac
arrest
Sibutramine and reporting of
cerebrovascular disease
Rosiglitazone and reporting of MI
Time Period ROR-1.96SE PRR-1.96SE IC-2SD ROR-1.96SE PRR-1.96SE IC-2SD ROR-1.96SE PRR-1.96SE IC-2SD
2005Q1 -1.72 -0.69 -2.83 -6.64 -2.16 -1.15 -1.86
2005Q2 -2.03 -1.03 -2.86 -4.89 -4.89 -3.83 -2.69 -1.6 -1.81
2005Q3 -0.87 0.13 -1.63 -4.38 -4.38 -2.7 -2.95 -1.84 -1.39
2005Q4 0.42 1.41 -0.27 -3.81 -3.81 -1.91 -2.87 -1.75 -1.13
2006Q1 0.4 1.39 -0.26 -3.74 -3.74 -1.71 -2.97 -1.83 -1.02
2006Q2 0.42 1.42 -0.18 -4.01 -4.01 -1.65 -2.79 -1.67 -0.87
2006Q3 0.41 1.4 -0.18 -3.97 -3.97 -1.52 -2.78 -1.78 -0.83
2006Q4 0.53 1.53 -0.05 -3.97 -3.97 -1.52 -2.64 -1.53 -0.76
2007Q1 0.66 1.65 0.06 -3.97 -3.97 -1.52 -2.53 -1.53 -0.73
2007Q2 0.63 1.62 0.04 -3.33 -3.14 -1.09 -1.71 -0.66 -0.43
2007Q3 0.65 1.64 0.05 -2.92 -2.79 -0.87 -1.25 -0.22 -0.27
2007Q4 0.64 1.63 0.05 -2.53 -2.53 -0.74 -0.99 0.06 -0.19
2008Q1 0.91 1.89 0.35 -2.42 -2.33 -0.68 -0.76 0.28 -0.13
2008Q2 0.94 1.92 0.38 -2.15 -2.08 -0.59 -0.39 0.63 -0.01
2008Q3 0.94 1.93 0.39 -2.08 -2.08 -0.58 -0.29 0.72 0.02
2008Q4 0.97 1.96 0.42 -0.32 -0.31 -0.02 0.18 1.17 0.21
2009Q1 1.26 2.23 0.78 -0.18 -0.17 0.05 0.64 1.6 0.48
2009Q2 1.25 2.22 0.77 -0.12 -0.11 0.07 0.73 1.69 0.55
2009Q3 1.24 2.22 0.77 0.1 0.1 0.18 1 1.94 0.78
2009Q4 1.26 2.24 0.79 0.31 0.3 0.28 1.17 2.1 0.96
2010Q1 1.4 2.37 1 0.4 0.39 0.34 1.27 2.19 1.07
2010Q2 1.38 2.35 0.97 0.58 0.56 0.45 1.4 2.31 1.24
2010Q3 1.4 2.37 1 0.77 0.75 0.6 1.59 2.48 1.54
2010Q4 1.43 2.4 1.04 0.86 0.83 0.66 1.69 2.57 1.7
ROR=Reporting Odds Ratio; PRR= Proportional Reporting Ratio; IC= Information Component; SD=Standard Deviation; SE= Standard Error. Bold Letters: Positive
Signal
Subeesh, et al.: A Comparative Study of Data Mining Algorithms
Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018 447
Table 6: The DEC detected by the three DMAs.
Sl. no Drug Event Combination ROR PRR IC
1 Propoxyphene and cardiac arrest
2 Sibutramine and cerebrovascular disorders X X
3 Rosiglitazone and MI
4 Sitagliptin and renal failure X X
5 Canagliozin and UTI X
Sensitivity 60% 40% 100%
‘√’ means that the DMA could identify the signals prior to the withdrawal date;
‘X’ means that algorithms did not identify the signals prior to the withdrawal date.
Table 7: Sensitivity of each DMA in terms of early detection.
Sl. No. Drug Event IDW IDD Status
change
date
Present status
ROR Earlier
Quarters
PRR Earlier
Quarters
IC Earlier
Quarters
1 Propoxyphene Cardiac Arrest 2010Q4 2009Q1 3 2009Q1 3 2007Q1 14 2010 Withdraw
from market
2 Sibutramine Cerebrovascular
disorders
2010Q4 ___ ___ ___ ___ 2009Q1 7 2010 Withdraw
from market
3 Rosiglitazone MI 2010Q4 2009Q3 5 2009Q4 4 2008Q3 9 2010 Withdraw
from market
4Sitagliptin Renal failure 2012Q4 ___ ___ ___ ___ 2011Q2 4 2012 Change in
label
5 Canagliozin Urinary Tract
Infection
2016Q2 2015Q4 1 ___ ___ 2015Q3 2 2016 Change in
label
Abbreviations: ROR = Reporting Odds Ratio; PRR = Proportional Reporting Ratio; IC = Information Component; IDW = Index Date of Withdrawn; IDD = Index Date of
Detection.
Table 5: Sitagliptin and reporting of renal failure and Canagliozin and
reporting of UTI.
Sitagliptin and reporting of
renal failure
Canagliozin and reporting
of UTI
Time
Period
ROR-
1.96SE
PRR-
1.96SE
IC-2SD ROR-
1.96SE
PRR-
1.96SE
IC-2SD
2006Q4 -0.13 0.89 -2.84 - - -
2007Q1 0.09 1.1 -1.22 - - -
2007Q2 -0.37 0.64 -0.74 - - -
2007Q3 -0.3 0.7 -0.49 - - -
2007Q4 -0.2 0.8 -0.34 - - -
2008Q1 -0.18 0.82 -0.28 - - -
2008Q2 -0.13 0.87 -0.27 - - -
2008Q3 -0.22 0.78 -0.26 - - -
2008Q4 -0.26 0.74 -0.23 - - -
2009Q1 -0.25 0.75 -0.2 - - -
2009Q2 -0.25 0.76 -0.17 - - -
2009Q3 -0.26 0.75 -0.16 - - -
2009Q4 -0.22 0.79 -0.13 - - -
2010Q1 -0.17 0.83 -0.1 - - -
2010Q2 -0.17 0.83 -0.08 - - -
2010Q3 -0.15 0.86 -0.07 - - -
2010Q4 -0.14 0.86 -0.06 - - -
2011Q1 -0.09 0.91 -0.03 - - -
2011Q2 -0.05 0.96 0 - - -
2011Q3 0.01 1.01 0.03 - - -
2011Q4 0 0.99 0.03 - - -
2012Q1 0 1 0.04 - - -
2012Q2 -0.01 0.99 0.04 - - -
2012Q3 -0.01 1 0.05 - - -
2012Q4 0.01 1.01 0.06 - - -
2013Q1 -0.01 0.99 0.05 - - -
2013Q2 -0.01 0.98 0.06 -1.39 -0.36 -4.4
2013Q3 -0.03 0.97 0.06 0.08 3.63 -1.81
2013Q4 -0.05 0.95 0.06 0.18 4.14 -1.04
2014Q1 -0.05 0.95 0.06 0.2 3.09 -0.87
2014Q2 -0.1 0.9 0.05 -0.21 0.8 -0.84
2014Q3 -0.16 0.85 0.04 -0.18 0.82 -0.62
2014Q4 -0.18 0.83 0.04 -0.06 0.96 -0.49
2015Q1 -0.17 0.83 0.04 0.41 1.39 -0.17
2015Q2 -0.22 0.78 0.03 -0.55 0.48 -0.15
2015Q3 - - - 0.9 1.83 0.53
2015Q4 - - - 1.11 1.97 0.76
ROR=Reporting Odds Ratio; PRR= Proportional Reporting Ratio; IC= Informa-
tion Component; SD=Standard Deviation; SE= Standard Error. Bold Letters: Posi-
tive Signal
Subeesh, et al.: A Comparative Study of Data Mining Algorithms
448 Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018
DECs identied as signals were considered as the measure of sensitivity.
Compared to the above two studies,11,15 the ndings of the present study
seemed to be more reliable because the reference standard considered
is more robust. Comparison of DMAs is possible in many ways, but the
main aspect which directly inuences the DMA values is the number of
reports. us, according to this study, a cumulative number of reports
should be considered as the measure of sensitivity. Early detection of
ADRs is the main advantage of DMAs,16-17 therefore it can be considered
as a measure of sensitivity.
is study indicates that IC is more sensitive in terms of early detection
as well as the number of reports. However, among the DMAs, the sensi-
tivity dierence is statistically not signicant. Moreover, the sensitivity of
DMAs may vary if a dierent database is used hence, we cannot conclude
that IC is the most sensitive DMA. Nevertheless, for a given set of data
and DMAs, IC showed more sensitivity than other DMAs. It has been
observed that the sensitivity is proportional to the number of reports.
According to V.G. Koutkias and M.-C. Jaulent,18 the number of reports is
an important factor for signal strength.
According to the literature review, it was observed that no study has been
attempted to compare DMA to identify the sensitive method among the
existing algorithms. We comply with P. Waller19 study, as the selection
of DMAs should be done on the basis of their specicity, sensitivity and
predictive value in addition to a factor observed from our study. Early
detection of ADR plays a vital role which may reduce the casualties and
provide sucient time for a regulatory decision.
It is not surprising that IDD is earlier than IDW because once the positive
signal is identied, FDA requires adequate time to evaluate the situation,
assess the risk-benet prole of the drug and for a regulatory decision.
However, the use of DMAs could trigger the initiation of this process
earlier by recognizing signals in advance. Few studies have compared the
traditional method of ADR detection with computer-based signal detec-
tion techniques. A study conducted by A.W. I20 concluded that DMAs
detected safety signals well before the conventional ways. According to
D.J. Graham,21 there were 88,000–140,000 excess of cardiac disorders
associated with rofecoxib. It could have been reduced if the safety signal
was detected earlier.
We rely on the progression between the early detection of an ADR signal
to the nal decision that drug withdrawal or labelling change could more
likely be acquired due to the earlier detection of ADR signals as a
result of the applications of DMAs. Consequently, the time when the
FDA makes a decision will correspondingly occur earlier.
Limitations of the study
ere are some concerns regarding FDA AERS database. Under-reporting
is the main concern with regards to any spontaneous reporting system
(SRS) database. SRS will not reect the actual picture of the scenario.
us, more oen the situation was underestimated. Secondly, the reporting
may get biased when there is a change in labelling criteria or any special
updates regarding an ADR of the drug. As a result, over-reporting of
that particular ADR will occur which will decrease the signal strength
of other ADRs of the same drug (change in Ni value will aect the signal
strength).
e selection of brand names for data mining is another limitation which
we had come across. FDA AERS database is a collection of ADR reports
around the world but the main contributor is United States (US). e
brand names used for data mining in this study mainly focused on the
US, European countries and India. us, the chances of missing data
cannot be ruled out.
CONCLUSION
e aim of the study was to compare and appraise the performance
of signal detection techniques used in data mining. It is the rst study
attempted to address the importance of early detection of ADR and iden-
tication of the sensitive method. Even though there is no statistically
signicant dierence among three DMAs, IC was found to be sensitive
method compared with other two DMAs in FDA AERs database. e
sensitivity and comparability of DMA is proportionate to the number of
reports of DEC.
CONFLICT OF INTEREST
e authors declare no conict of interest.
ABBREVIATIONS
FDA: Food and Drug Administration; AERS: Adverse Events Report-
ing System: AEs: Adverse Event; PEM: Prescription-Event Monitoring;
DMAs: Data Mining Algorithms; ADRs: Adverse Drug Reactions; ROR
: Reporting Odds Ratio; MGPS: Multi-item Gamma Poisson Shrinker;
PRR: Proportional Reporting Ratio; IC: Information Component;
DEC: Drug Event Combinations; IDW/L: Index Date of Withdrawn/
Label change.
REFERENCES
1. Harpaz R, DuMouchel W, LePendu P, Bauer MA, Ryan P, Shah NH. Performance
of pharmacovigilance signal-detection algorithms for the fda adverse event
reporting system. Clin Pharmacol Ther. 2013;93(6):539-46. 10.1038/clpt.2013.24
2. Bie SD, Ferrajolo C, Straus SM, Verhamme KM, Bonhoeffer J, Wong IC, et al.
Pediatric drug safety surveillance in fda-aers: A description of adverse
events from grip project. PLoS One. 2015;10(6):e0130399. 10.1371/journal.
pone.0130399
3. Rossi AC, Knapp DE, Anello C, et al. Discovery of adverse drug reactions: A
comparison of selected phase iv studies with spontaneous reporting methods.
JAMA. 1983;249(16):2226-8. 10.1001/jama.1983.03330400072029
4. Rothman KJ, Lanes S, Sacks ST. The reporting odds ratio and its advantages
over the proportional reporting ratio. Pharmacoepidemiology and Drug Saf.
2004;13(8):519-23. 10.1002/pds.1001
5. Napoli AA, Wood JJ, Coumbis JJ, Soitkar AM, Seekins DW, Tilson HH. No
evident association between efavirenz use and suicidality was identied
from a disproportionality analysis using the faers database. J Int AIDS Soc.
2014;17(1):9214. 10.7448/ias.17.1.19214
6. Evans SJ, Waller PC, Davis S. Use of proportional reporting ratios (prrs) for
signal generation from spontaneous adverse drug reaction reports. Pharmaco-
epidemiol Drug Saf. 2001;10(6):483-6. 10.1002/pds.677
7. Bate A, Evans SJ. Quantitative signal detection using spontaneous adr reporting.
Pharmacoepidemiol Drug Saf. 2009;18(6):427-36. 10.10 02/pds.1742
8. Bate A, Lindquist M, Edwards IR, Orre R. A data mining approach for signal
detection and analysis. Drug Saf. 2002;25(6):393-7.
9. Wilson AM, Thabane L, Holbrook A. Application of data mining techniques in
Figure 1: Sensitivity associated with the number of reports per DEC.
Abbreviations: ROR = Reporting Odds Ratio; PRR = Proportional Reporting
Ratio; IC = Information Component.
Subeesh, et al.: A Comparative Study of Data Mining Algorithms
Journal of Young Pharmacists, Vol 10, Issue 4, Oct-Dec, 2018 449
Indian J Pharmacol. 2015; 47(3):241-2. 10.4103/0253-7613.157102
1 7. Hauben M, Reich L. Potential utility of data-mining algorithms for early detection
of potentially fatal/disabling adverse drug reactions: A retrospective evaluation.
J Clin Pharmacol. 2005;45(4):378-84. 10.1177/0091270004273936
18. Koutkias VG, Jaulent MC. Computational approaches for pharmacovigilance
signal detection: Toward integrated and semantically-enriched frameworks.
Drug Saf. 2015;38(3):219-32. 10.10 07/s40264-015-0278-8
19. Waller P, Puijenbroek EV, Egberts A, Evans S. The reporting odds ratio versus
the proportional reporting ratio:‘Deuce’. Pharmacoepid and drug saf. 2004;
13(8):525-6.
20. I AW, Pratt NL, Kalisch LM, Roughead EE. Comparing time to adverse drug
reaction signals in a spontaneous reporting database and a claims database: A
case study of rofecoxib-induced myocardial infarction and rosiglitazone-induced
heart failure signals in australia. Drug Saf. 2014;37(1):53-64. 10.1007/s40264-
013-0124-9
21. Graham DJ, Campen D, Hui R, Spence M, Cheetham C, Levy G, et al. Risk of
acute myocardial infarction and sudden cardiac death in patients treated with
cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inammatory
drugs: Nested case-control study. Lancet. 2005;365(9458):475-81. 10.1016/
s0140-6736(05)17864-7
pharmacovigilance. Br J Clin Pharmacol. 2004;57(2):127-34. 10.1046/j.1365-
2125.2003.01968.x
10. Szarfman A, Machado SG, O’Neill RT. Use of screening algorithms and computer
systems to efciently signal higher-than-expected combinations of drugs and
events in the us fda’s spontaneous reports database. Drug Saf. 2002;25(6):381-92.
11. Puijenbroek EPV, Bate A, Leufkens HG, Lindquist M, Orre R, Egberts AC.
A comparison of measures of disproportionality for signal detection in sponta-
neous reporting systems for adverse drug reactions. Pharmacoepidemiol Drug
Saf. 2002;11(1):3-10. 10.1002/pds.668
12. Shamloo B, Chhabra P, Freedman A, Potosky A, Malin J, Smith S. Novel adverse
events of bevacizumab in the us fda adverse event reporting system database.
Drug Saf. 2012;35(6):507-18. 10.2165/11597600-000000000-000 00
13. Deshpande G, Gogolak V, Smith S. Data mining in drug safety. Pharmaceut
Med. 2010;24(1):37-43. 10.1007/BF03256796
14. Lindquist M, Stahl M, Bate A, Edwards IR, Meyboom RH. A retrospective evalu-
ation of a data mining approach to aid nding new adverse drug reaction signals
in the who international database. Drug Saf. 2000;23(6):533-42.
15. Kubota K, Koide D, Hirai T. Comparison of data mining methodologies using
japanese spontaneous reports. Pharmacoepidemiol Drug Saf. 2004;13(6):387-94.
10.10 02/pds.964
16. Chakraborty BS. Pharmacovigilance: A data mining approach to signal detection.
Article History: Submission Date : 19-05-2018; Revised Date : 30-06-2018; Acceptance Date : 06-08-2018.
Cite this article: Subeesh V, Maheswari E, Saraswathy GR, Swaroop AM, Minnikanti SS. A Comparative Study of Data Mining Algorithms used for Signal
Detection in FDA AERS Database. J Young Pharm. 2018;10(4):444-9.
... Pharmacovigilance signal measures, including the reporting odds ratio (ROR), proportional reporting ratio (PRR), information component (IC), and empirical Bayesian geometric mean (EBGM), were calculated in each dataset, using the R package PhViD [75]. All four pharmacovigilance measures were calculated due to differences in their sensitivity and early detection potential [76,77]; for brevity, only the PRR is shown in the text, and all calculated measures can be found in the Supplementary Materials. Given the support for the use of the false discovery rate (FDR) to identify signals over thresholds, we used an FDR < 0.05 to denote significance [78,79]. ...
... Appendix SA: Materials and Methods. References [32,63,[73][74][75][76][77][78][79][84][85][86][87] are cited in the supplementary materials. protection of privacy and integrity of individuals was guaranteed, and in order to safeguard the identity of individuals certain data elements, names/identifiers or country-specific information were not disclosed by the EMA [28]. ...
Article
Full-text available
Abstract: In the past twenty years, the consumption of opioid medications has reached significant proportions, leading to a rise in drug misuse and abuse and increased opioid dependence and related fatalities. Thus, the purpose of this study was to determine whether there are pharmacovigilance signals of abuse, misuse, and dependence and their nature for the following prescription opioids: codeine, dihydrocodeine, fentanyl, oxycodone, pentazocine, and tramadol. Both the pharmacovigilance datasets EudraVigilance (EV) and the FDA Adverse Events Reporting System (FAERS) were analyzed to identify and describe possible misuse-/abuse-/dependence-related issues. A descriptive analysis of the selected Adverse Drug Reactions (ADRs) was performed, and pharmacovigilance signal measures (i.e., reporting odds ratio, proportional reporting ratio, information component, and empirical Bayesian geometric mean) were computed for preferred terms (PTs) of abuse, misuse, dependence, and withdrawal, as well as PTs eventually related to them (e.g., aggression). From 2003 to 2018, there was an increase in ADR reports for the selected opioids in both datasets. Overall, 16,506 and 130,293 individual ADRs for the selected opioids were submitted to EV and FAERS, respectively. Compared with other opioids, abuse concerns were mostly recorded in relation to fentanyl and oxycodone, while tramadol and oxycodone were more strongly associated with drug dependence and withdrawal. Benzodiazepines, antidepressants, other opioids, antihistamines, recreational drugs (e.g., cocaine and alcohol), and several new psychoactive substances, including mitragynine and cathinones, were the most commonly reported concomitant drugs. ADRs reports in pharmacovigilance databases confirmed the availability of data on the abuse and dependence of prescription opioids and should be considered a resource for monitoring and preventing such issues. Psychiatrists and clinicians prescribing opioids should be aware of their misuse and dependence liability and effects that may accompany their use, especially together with concomitant drugs.
... In general, document frequency is the simplest and least computational feature selection method. e information gain of a word is a measure of the information increment brought by the word to the classification, which reflects the ability of the word to distinguish this category from other categories [25]. It performs category prediction by calculating the number of instances in which a word is included and the number of instances in which it is not included. ...
Article
Full-text available
English correspondence writing has become a necessary skill for every scientific researcher and high-tech talents. An English correspondence writing auxiliary writing system can help nonnative English speakers make up for the lack of professional expression. The key factor of business English correspondence writing system is the construction of knowledge base. To improve the business English correspondence writing knowledge base, we need to mine frequent patterns of sentences in each category. The purpose of this topic is to improve and supplement the knowledge base for the business English correspondence writing system and propose frequent pattern mining for sentences in each category, so as to improve the writing knowledge base for the business English correspondence writing system. Firstly, we crawl a large number of business English letters and telegrams from the Internet, extract the relevant summary information, then store it, and preliminarily construct a corpus based on sentences. Then, we do some research on the structure of business English correspondence abstracts, mark the sentences in the corpus and count the relevant information, and have a certain understanding of their writing methods. Finally, we mine frequent patterns for sentences in each category, so as to improve the knowledge base of summary writing for the business English correspondence writing system. In the experiment, we use the classical FP growth algorithm as the mining method. The experiment shows that the frequent patterns between 3 and 6 words have been mined to a certain extent. By gradually improving the mining strategy, the quality of mining results has been improved and the writing effect of business English correspondence of scientific researchers has been improved.
... Pharmacovigilance signal measures, including the reporting odds ratio (ROR), proportional reporting ratio (PRR), information component (IC), and empirical Bayesian geometric mean (EBGM), were calculated in each dataset using the R ® package PhViD [102]. All four pharmacovigilance measures were calculated due to differences in their sensitivity and early detection potential [18,101,[103][104][105][106]; for brevity, only the PRR is shown in the text; all calculated measures can be found in the supplemental tables. Signals are dispro-portionality measures based on a 2 × 2 contingency table; they help determine whether a drug adverse event pair occurs more often than expected by comparing signal values to published thresholds [107,108]. ...
Article
Full-text available
Despite increasing reports, antidepressant (AD) misuse and dependence remain underestimated issues, possibly due to limited epidemiological and pharmacovigilance evidence. Thus, here we aimed to determine available pharmacovigilance misuse/abuse/dependence/withdrawal signals relating to the Selective Serotonin Reuptake Inhibitors (SSRI) citalopram, escitalopram, paroxetine, fluoxetine, and sertraline. Both EudraVigilance (EV) and Food and Drug Administration-FDA Adverse Events Reporting System (FAERS) datasets were analysed to identify AD misuse/abuse/dependence/withdrawal issues. A descriptive analysis was performed; moreover, pharmacovigilance measures, including the reporting odds ratio (ROR), the proportional reporting ratio (PRR), the information component (IC), and the empirical Bayesian geometric mean (EBGM) were calculated. Both datasets showed increasing trends of yearly reporting and similar signals regarding abuse and dependence. From the EV, a total of 5335 individual ADR reports were analysed, of which 30% corresponded to paroxetine (n = 1,592), 27% citalopram (n = 1,419), 22% sertraline (n = 1,149), 14% fluoxetine (n = 771), and 8% escitalopram (n = 404). From FAERS, a total of 144,395 individual ADR reports were analysed, of which 27% were related to paroxetine, 27% sertraline, 18% citalopram, 16% fluoxetine, and 13% escitalopram. Comparing SSRIs, the EV misuse/abuse-related ADRs were mostly recorded for citalopram, fluoxetine, and sertraline; conversely, dependence was mostly associated with paroxetine, and withdrawal to escitalopram. Similarly, in the FAERS dataset, dependence/withdrawal-related signals were more frequently reported for paroxetine. Although SSRIs are considered non-addictive pharmacological agents, a range of proper withdrawal symptoms can occur well after discontinuation, especially with paroxetine. Prescribers should be aware of the potential for dependence and withdrawal associated with SSRIs.
... In this example, the dimensionality of the dataset is 70. Different disproportionality methods have been considered for finding binary associations in a number of works [19][20][21]. But in case of multidimensional databases, these methods require subgroup analysis and to consider each and every combination between different attributes based on their values. ...
Article
The activity of post-marketing surveillance results in a collection of large amount of data. The analysis of data is very useful for raising early warnings on possible adverse reactions of drugs. Association rule mining techniques have been heavily explored by the research community for identifying binary association between drugs and their adverse effects. But these techniques perform poorly and miss out several interesting associations when it comes to analysis of multidimensional data which may include multiple patient attributes, drugs and adverse drug reactions. In the present work, a clustering-based hybrid approach has been presented for finding quantitative multidimensional association from the large amount of data. Firstly, it employs clustering technique for segmentation of data into semantically coherent clusters. Furthermore, disproportionality method called proportional reporting ratio is applied on clustered data for generating statistically strong associations. The performance of the proposed methodology has been examined on the data taken from the U.S. Food and Drug Administration Adverse Event Reporting System database corresponding to Aspirin and nine other drugs which are prescribed along with Aspirin. The experimental results show that the proposed approach discovered a number of association rules which are very comprehensive and informative regarding relationship of patient traits and drugs with adverse drug reactions. On comparing experimental results with LPMiner, it is observed that the quantitative association rules discovered by LPMiner are just 8.3% of what have been discovered by the proposed methodology.
... In such contexts, data mining techniques have been deployed as fast and efficient ways to detect possible ADR signals coming from different sources of evidence. Several data mining algorithms have been described in literature [41], mainly based on nonparametrical statistics theory, including the Reporting Odds Ratio (ROR), Multi-item Gamma Poisson Shrinker (MGPS), the Proportional Reporting Ratio (PRR), and the Information Component (IC) [42] to name a few examples. An overview of statistical signal detection methods has been outlined in [43], whereas the first seminal examination of data mining techniques was given by Hauben over several articles [44][45][46]. ...
Article
Full-text available
Today’s surge of big data coming from multiple sources is raising the stakes that pharmacovigilance has to win, making evidence synthesis a more and more robust approach in the field. In this scenario, many scholars believe that new computational methods derived from data mining will effectively enhance the detection of early warning signals for adverse drug reactions, solving the gauntlets that post-marketing surveillance requires. This article highlights the need for a philosophical approach in order to fully realize a pharmacovigilance 2.0 revolution. A state of the art on evidence synthesis is presented, followed by the illustration of E-Synthesis, a Bayesian framework for causal assessment. Computational results regarding dose-response evidence are shown at the end of this article.
Article
Background Efficiency and accuracy for signal detection and evaluation activities are integral components of routine pharmacovigilance (PV) practices. However, an individual case safety report (ICSR) may consist of a variety of confounders such as concomitant medications (CM), past medical history (PMH), and concurrent medical conditions that influence a safety officer’s evaluation of a potential adverse event (AE). Limited pharmacovigilance systems are currently available as a tool designed to enhance the efficiency and accuracy of signal detection and management. Objective To introduce a systemic approach to make critical safety information readily available for users in order to discern possible interferences from CM and make informed decisions on the signal evaluation process – saving time while improving quality. Methods Oracle Empirica Signal software was utilized to extract cases with CM that are Known Implicating Medications (KIM) for each AE according to public regulatory information from drug labels – FDA Structured Product Labeling (SPL) or EMA Summary of Product Characteristics (SPC). SAS Enterprise Guide was used to further process the data generated from Oracle Empirica Signal software. Results For any target drug being evaluated for safety purposes, a KIM reference table can be generated which summarizes all potential causality contributions from CMs. Conclusion In addition to providing standalone KIM table as reference, adoption of this concept and automation may also be fully integrated into commercial signal detection and management software packages for easy use and accessibility and may even lead to reduced False Positive rate in signal detection within the PV space.
Article
Full-text available
The detection of new drug safety signals is of growing importance with ever more new drugs becoming available and exposure to medicines increasing. The task of evaluating information relating to safety lies with national agencies and, for international data, with the World Health Organization Programme for International Drug Monitoring. An established approach for identifying new drug safety signals from the international database of more than 2 million case reports depends upon clinical experts from around the world. With a very large amount of information to evaluate, such an approach is open to human error. To aid the clinical review, we have developed a new signalling process using Bayesian logic, applied to data mining, within a confidence propagation neural network (Bayesian Confidence Propagation Neural Network; BCPNN). Ultimately, this will also allow the evaluation of complex variables. The first part of this study tested the predictive value of the BCPNN in new signal detection as compared with reference literature sources (Martindale's Extra Pharmacopoeia in 1993 and July 2000, and the Physicians Desk Reference in July 2000). In the second part of the study, results with the BCPNN method were compared with those of the former signalling procedure. In the study period (the first quarter of 1993) 107 drug-adverse reaction combinations were highlighted as new positive associations by the BCPNN, and referred to new drugs. 15 drug-adverse reaction combinations on new drugs became negative BCPNN associations in the study period. The BCPNN method detected signals with a positive predictive value of 44% and the negative predictive value was 85%. 17 as yet unconfirmed positive associations could not be dismissed with certainty as false positive signals. Of the 10 drug-adverse reaction signals produced by the former signal detection system from data sent out for review during the study period, 6 were also identified by the BCPNN. These 6 associations have all had a more than 10-fold increase of reports and 4 of them have been included in the reference sources. The remaining 4 signals that were not identified by the BCPNN had a small, or no, increase in the number of reports, and are not listed in the reference sources. Our evaluation showed that the BCPNN approach had a high and promising predictive value in identifying early signals of new adverse drug reactions.
Article
Computational signal detection constitutes a key element of postmarketing drug monitoring and surveillance. Diverse data sources are considered within the 'search space' of pharmacovigilance scientists, and respective data analysis methods are employed, all with their qualities and shortcomings, towards more timely and accurate signal detection. Recent systematic comparative studies highlighted not only event-based and data-source-based differential performance across methods but also their complementarity. These findings reinforce the arguments for exploiting all possible information sources for drug safety and the parallel use of multiple signal detection methods. Combinatorial signal detection has been pursued in few studies up to now, employing a rather limited number of methods and data sources but illustrating well-promising outcomes. However, the large-scale realization of this approach requires systematic frameworks to address the challenges of the concurrent analysis setting. In this paper, we argue that semantic technologies provide the means to address some of these challenges, and we particularly highlight their contribution in (a) annotating data sources and analysis methods with quality attributes to facilitate their selection given the analysis scope; (b) consistently defining study parameters such as health outcomes and drugs of interest, and providing guidance for study setup; (c) expressing analysis outcomes in a common format enabling data sharing and systematic comparisons; and (d) assessing/supporting the novelty of the aggregated outcomes through access to reference knowledge sources related to drug safety. A semantically-enriched framework can facilitate seamless access and use of different data sources and computational methods in an integrated fashion, bringing a new perspective for large-scale, knowledge-intensive signal detection.
Article
Data mining is used in pharmacovigilance as an adjunct to traditional pharmacovigilance practices. There remains ongoing debate as to the impact automated signal detection would have on pharmacovigilance resources. An important component of this debate is the value of each statistical alert or signal of disproportional reporting (SDR) and the resources needed to evaluate SDRs that are clinically unimportant. Using the terminology of diagnostic testing, such SDRs are called false positives as they are statistically positive but are clinically negative. Based on the clinical testing paradigm, a more stringent threshold increases the sensitivity of the test by lowering the number of false positives; however, the trade off of increased sensitivity is a reduced specificity, i.e. potentially missing clinically relevant problems. In developing the protocol to assess the clinical validity of an SDR, a literature search was conducted to determine what threshold(s) were commonly used for data mining adverse event databases. Of the more than 100 manuscripts identified, 41 published the results of data mining excursions with a clearly identified threshold for significance. The commonly used data mining algorithms were proportional reporting ratio (PRR), reporting odds ratio (ROR), multi-item gamma Poisson shrinker (MGPS) and Bayesian confidence propagation neural network (BPCNN). There was some variation in the threshold used for each algorithm. For the PRR, thresholds of 1.0, 1.5 and 2.0 were reported. Some authors required a Chi-squared test statistic of ≥4.0. Minimum drug-event pair counts were most often required for the frequentist measures of disproportionality, PRR and ROR. Among the Bayesian algorithms, MGPS and BPCNN, there was variation in the metric used and, within a metric, variation in thresholds and the use of minimum case counts. Metrics based on the MGPS algorithm that have been used to determine statistical significance include the empirical Bayes geometric mean (EBGM), the lower 95% confidence interval of the EBGM, the lower 95% confidence interval of the empirical Bayes arithmetic mean, EBlog2 and interaction signal score. Based on the published literature, there is considerable variation in defining a significant alert or SDR among practitioners of pharmacovigilance data mining. Research into the impact of such variations in practice on SDR volume and value is urgently needed.
Article
Quantitative methods are increasingly used to analyse spontaneous reports. We describe the core concepts behind the most common methods, the proportional reporting ratio (PRR), reporting odds ratio (ROR), information component (IC) and empirical Bayes geometric mean (EBGM). We discuss the role of Bayesian shrinkage in screening spontaneous reports, the importance of changes over time in screening the properties of the measures. Additionally we discuss three major areas of controversy and ongoing research: stratification, method evaluation and implementation. Finally we give some suggestions as to where emerging research is likely to lead. Copyright © 2009 John Wiley & Sons, Ltd.
Article
Bevacizumab is the first in its class, vascular endothelial growth factor (VEGF) inhibitor that was initially approved by the US FDA in 2004 for the treatment of metastatic colon cancer and other solid tumors. Preapproval clinical trials, particularly for oncology drugs, are limited in their ability to detect certain adverse effects and, therefore, the FDA and pharmaceutical sponsors collect and monitor reports of adverse events (AEs) following approval. The purpose of this study was to screen the FDA's Adverse Event Reporting System (AERS) database for novel AEs that may be attributed to bevacizumab. The FDA AERS database was used to identify all AE reports for bevacizumab from February 2004 to September 2009. Disproportionality analysis was conducted for bevacizumab against all other drugs in the background by setting statistical significance at proportional reporting ratio (PRR) ≥2, observed case count ≥3 and chi-square ≥4. Subsequent clinical evaluation was performed to determine the clinical relevance of the findings and to group related events. A total of 523 Preferred Terms (PTs) were disproportionally reported; following clinical review 63 (12%) were found to be both unlabelled and of clinical importance. These PTs were grouped into 15 clinical disorder groups. Among the clinical disorders, electrolyte abnormalities had the greatest number of reports (n = 426) followed by cardiovascular events (n = 421), gastrointestinal events (n = 345), nervous system disorders (n = 106) and pneumonitis (n = 96). On sensitivity analysis, a number of clinically important unlabelled disorders, such as necrotizing fasciitis, vessel wall disorders, arrhythmia and conduction disorder and autoimmune thrombocytopenia still met the statistical significance criteria. During the study period, out of 12 010 AE reports mentioning bevacizumab, it was listed as the suspect drug in 94.2% of the reports. Our disproportionality analysis identified many events that are already recognized as AEs of bevacizumab, but it also identified a number of clinically important unlabelled terms, which if confirmed in future studies would have potential implications for use of bevacizumab in clinical practice.
Article
The process of generating 'signals' of possible unrecognized hazards from spontaneous adverse drug reaction reporting data has been likened to looking for a needle in a haystack. However, statistical approaches to the data have been under-utilised. Using the UK Yellow Card database, we have developed and evaluated a statistical aid to signal generation called a Proportional Reporting Ratio (PRR). The proportion of all reactions to a drug which are for a particular medical condition of interest is compared to the same proportion for all drugs in the database, in a 2 x 2 table. We investigated a group of newly-marketed drugs using as minimum criteria for a signal, 3 or more cases, PRR at least 2, chi-squared of at least 4. The database was used to examine retrospectively 15 drugs newly-marketed in the UK, with the highest levels of ADR reporting. The method identified 481 signals meeting the minimum criteria during the period 1996-8. Further evaluation of these showed that 70% were known adverse reactions, 13% were events which were likely to be related to the underlying disease and 17% were signals requiring further evaluation. Proportional reporting ratios are a valuable aid to signal generation from spontaneous reporting data which are easy to calculate and interpret, and various refinements are possible.