Next Article in Journal
Exploring Polymeric Micelles for Improved Delivery of Anticancer Agents: Recent Developments in Preclinical Studies
Previous Article in Journal / Special Issue
Vaccine Safety Surveillance Systems: Critical Elements and Lessons Learned in the Development of the US Vaccine Safety Datalink’s Rapid Cycle Analysis Capabilities
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:

Drug Adverse Event Detection in Health Plan Data Using the Gamma Poisson Shrinker and Comparison to the Tree-based Scan Statistic

Department of Population Medicine, Harvard Medical School and Harvard Pilgrim Health Care Institute, 133 Brookline Avenue, 6th Floor, Boston, MA 02215, USA
Pfizer, Inc., New York, NY 10017, USA
Kaiser Permanente Georgia, Atlanta, GA 30305, USA
OptumInsight, Waltham, MA 02451, USA
Harvard School of Public Health, Boston, MA 02115, USA
Meyers Primary Care Institute, University of Massachusetts Medical School, the Meyers Primary Care Institute, Fallon Community Health Plan, Worcester, MA 01605, USA
Center for Health Studies, Group Health Cooperative, Seattle, WA 98101, USA
Lovelace Clinic Foundation, Albuquerque, NM 87106, USA
Kaiser Permanente Northern California, Oakland, CA 94611, USA
HealthPartners Research Foundation, Minneapolis, MN 55440, USA
Kaiser Permanente Colorado, Denver, CO 80237, USA
Kaiser Permanente Northwest, Portland OR 97227, USA
Author to whom correspondence should be addressed.
Pharmaceutics 2013, 5(1), 179-200;
Received: 6 November 2012 / Revised: 1 March 2013 / Accepted: 4 March 2013 / Published: 14 March 2013
(This article belongs to the Special Issue Drug Safety and Pharmacovigilance)


Background: Drug adverse event (AE) signal detection using the Gamma Poisson Shrinker (GPS) is commonly applied in spontaneous reporting. AE signal detection using large observational health plan databases can expand medication safety surveillance. Methods: Using data from nine health plans, we conducted a pilot study to evaluate the implementation and findings of the GPS approach for two antifungal drugs, terbinafine and itraconazole, and two diabetes drugs, pioglitazone and rosiglitazone. We evaluated 1676 diagnosis codes grouped into 183 different clinical concepts and four levels of granularity. Several signaling thresholds were assessed. GPS results were compared to findings from a companion study using the identical analytic dataset but an alternative statistical method—the tree-based scan statistic (TreeScan). Results: We identified 71 statistical signals across two signaling thresholds and two methods, including closely-related signals of overlapping diagnosis definitions. Initial review found that most signals represented known adverse drug reactions or confounding. About 31% of signals met the highest signaling threshold. Conclusions: The GPS method was successfully applied to observational health plan data in a distributed data environment as a drug safety data mining method. There was substantial concordance between the GPS and TreeScan approaches. Key method implementation decisions relate to defining exposures and outcomes and informed choice of signaling thresholds.

1. Introduction

Quantitative identification of unspecified medical product-adverse event (AE) relationships—often referred to as signal detection—is integral to worldwide medical product safety surveillance. Gamma Poisson Shrinkage (GPS) is a disproportionality method commonly applied to spontaneous reporting systems for signal detection [1]. Implementation of signal detection methods using routinely collected electronic data can expand the scope and scale of pharmacovigilance. In contrast with spontaneous reporting systems, however, little experience has been gained in the implementation and interpretation of GPS with observational electronic health care claims and administrative data.
Investigators have proposed a variety of AE signal detection methods for observational data, including disproportionality approaches [1,2,3,4], the tree-based scan statistic (TreeScan) [5,6] and others [7,8,9,10]. Disproportionality approaches, including GPS, Information Component (IC) and the proportional reporting ratio (PRR) all have been applied to observational data, typically in two fundamentally different ways. One approach has been to apply the methods as closely as possible to their implementation in spontaneous report datasets by using observational data to mimic spontaneous reports of drug-event combinations [2,11], including the “spontaneous reporting system methods” of GPS, Information Component, PRR and reporting odds ratio (ROR) as implemented and evaluated in Schuemie et al. (2012). In one example, Curtis et al. (2008) identified exposure using the Medicare Current Beneficiary Survey (MCBS) and outcomes from a sample of medical claims linked to the MCBS. Monthly reports were created to mimic spontaneous reporting databases and analyzed as if they were spontaneous reports. Zorych (2011) used simulated and administrative claims data to evaluate disproportionality methods using three different approaches for creating the analytic 2 × 2 table; none accounted for exposed or unexposed person time. Schuemie (2011) used simulated data to conduct a pilot implementation of several modifications of GPS, comparing person-level and exposure-day level approaches for calculating observed and expected counts, and specifically adjusting for protopathic bias [3].
A second approach adapts these methods to try to better leverage the richness of longitudinal observational datasets. Noren et al. (2008; 2010) applied the Information Component Temporal Pattern Discovery (ICTPD) approach by comparing the observed count of a drug-outcome combination to an expected count based on general occurrences in the database, coupled with a self-controlled design element by comparing the Observed and Expected counts of an event after prescription to the Observed and Expected counts before prescription [4,12]. Schuemie et al. (2011) used simulated data to evaluate an alternative approach (Longitudinal GPS: LGPS) similar to our implementation here where rather than comparing to expected counts based on occurrence of events for patients taking other prescribed products he utilizes exposed and non-exposed time at risk to develop a richer denominator [13].
Both implementation approaches have strengths and weaknesses. The LGPS method computes expected counts of medical events during drug exposure based on an aggregate of unexposed patient time in ever‐exposed and unexposed patients, potentially introducing confounding as unexposed patients may be less likely to have events related to the drug indication or underlying disease than the exposed population. For ICTPD, one of the two comparisons is of events occurring within a specific time after a dispensing of the drug of interest to all observations of that event after exposure to all other drugs but within the same at-risk period to give an Expected count. Inclusion of drugs associated with the outcome of interest will inflate the Expected count, and could lead to a reduced ICTPD score for the drug-event of interest; the inverse could occur with protective effects [14,15,16]. The GPS and ICTPD approaches and others differ in how a score is derived for the drug-outcome pairs, but also in terms of the test statistic, the choice of signaling threshold, as well as differences in implementation, some of which reflect the differences in the observational databases used (e.g., different terminological classifications of outcomes) [17].
More recently, Schuemie et al. (2012) and Ryan et al. (2012) published a comparison of multiple signal detection methods using longitudinal data across three countries [13,18]. The approaches are similar. Schuemie (2012) compared 10 methods using a set of positive and negative controls (drug-event pairs) for comparison. They reported positive results for most methods, including LGPS. Direct applicability of their results to routine open ended signal detection is hard to assess as they limited their assessment to a small set of known associations and their comparisons were based on area under the curve estimates on ROC curves where all sensitivity and specificity thresholds are considered equally important. In practice the tail ends of ROC curves may not be appropriate to consider in assessing surveillance approach effectiveness; if there is great differential performance between methodological approaches in these tails of ROC curves a misleading impression of performance and erroneous comparisons between approaches can be made. Schuemie et al. (2012) focused on point estimates instead of the lower thresholds of confidence limits that are more commonly used in signal detection to protect against spurious findings [1]. Finally, focusing on point estimates creates the potential to favor methods that routinely over-estimate risk.
Given that the GPS approach has shown promising results for use in longitudinal data [2,3,5,11,13], we furthered the prior work by applying GPS in a “real-world” environment not limited to specific associations but rather including non-prespecified drug-event pairs for evaluation. Such a real-world open ended discovery approach has not to our knowledge been taken with a GPS based method, although open ended discovery was done in Noren et al. (2010) for the ICTPD approach [12]. Our implementation closely mimicked the approach described in the U.S. Food and Drug Administration’s (FDA) Mini-Sentinel project for evaluation of non-prespecified AEs [19].
We present a pilot study evaluating the implementation of GPS for drug-AE signal detection using routinely-collected electronic medical encounter data in a multi-site environment. We also compare the GPS results to findings from a TreeScan study that used identical input datasets.

2. Methods

2.1. Overview

Signal detection using observational data requires three key specifications: (i) the analytic approach related to calculating exposures, identifying cases, defining comparators, and handling censoring; (ii) the statistical method used; and (iii) the signaling thresholds. Our implementation compared the rate of exposed outcomes with an expected count based on unexposed time. Therefore, the specific question was whether there is a statistical signal of excess risk of an outcome during exposed time as compared to unexposed time. In this paper we define a “signal” as a statistical association between a drug and a diagnosis within an exploratory framework without any requirement for verification of case status by medical record review or other confirmatory analysis. These “statistical signals” do not imply causality, but rather represent an association that meets pre-specified signaling thresholds that may warrant further investigation. Statistical signals identified using signal detection methods often can be explained by bias and confounding. We focused on signal detection implementation approaches using observational data, not prioritization and investigation of the signals identified. Such signal refinement requires additional methods and dedicated resources beyond the scope of this study [20].

2.2. Data and Study Population

The study cohort consisted of approximately 3.4 million privately-insured health plan members enrolled between 1999 and 2003 distributed roughly equally across the nine plans in the HMO Research Network Center for Education and Research on Therapeutics: Harvard Pilgrim Health Care, Kaiser Permanente Georgia, Meyers Primary Care Institute, Group Health Cooperative, Lovelace Clinic Foundation, Kaiser Permanente Northern California, HealthPartners Research Foundation, Kaiser Permanente Colorado, and Kaiser Permanente Northwest. Each health plan maintains an electronic database of member demographics, enrollment, outpatient pharmacy dispensing, and inpatient and outpatient encounters. These data have been used in several drug safety studies [21,22,23,24,25,26,27] and described in detail elsewhere [21,22,28].
Demographic information includes date of birth and sex. Enrollment consists of enrollment start and stop dates and a drug coverage indicator. Pharmacy dispensing data includes dispensing date, national drug code, units dispensed, and days supplied. Encounter information includes all diagnosis codes recorded during ambulatory and inpatient encounters.
We employed a distributed data model [29,30,31] that enabled sites to share only summarized count information for aggregation and analysis. The study was approved by the Institutional Review Board at each site.

2.3. Study Drugs

We identified users of two antifungal drugs, terbinafine and itraconazole, and two diabetes drugs, pioglitazone and rosiglitazone. These products were selected because they have substantial exposure, well-characterized risks, and allow for within-indication comparisons. We noted established associations between terbinafine and itraconazole and risk of liver disease [32,33] and allergic reaction [34,35]. Itraconazole and both diabetes drugs carry black box warnings for congestive heart failure on the U.S. FDA approved product labeling . Each drug was analyzed separately, without consideration of prior or concurrent exposures.

2.4. Diagnosis Definitions

Starting with all ICD-9-CM diagnosis codes we removed diagnosis codes associated with conditions unlikely to be drug-associated acute AEs (e.g., neoplasms, pregnancy and perinatal conditions, congenital anomalies, injuries and poisoning, diabetes). The remaining 1676 diagnosis codes were grouped using the Multi-level Clinical Classifications Software (MLCCS) [36]. The MLCCS is a hierarchical system with four levels of clinical concepts denoted by four 2-digit identifiers. The top level MLCCS identifies 18 body systems, and each can have up to three sublevels, as represented by the 2nd, 3rd, and 4th 2-digit codes. Each diagnosis code belongs to one classification group at each level of the MLCCS system, creating a hierarchical tree structure, where related diagnoses are close to each other on the tree. The exclusion process resulted in 183 overlapping groupings of related clinical concepts that were evaluated as potential AEs. Table 1 illustrates the hierarchical tree structure. Analyses were done at all four levels of granularity separately. Since we created a single set of diagnoses across products, we expected to identify some “signals” that represent bias and confounding common to uncontrolled observational studies (e.g., pioglitazone patients will have a higher rate of diabetes related complications such as eye disorders).

2.5. Contributed Person Time

All individuals with a membership period with medical and drug coverage over 180 days contributed person time. Membership gaps of 60 days or less were bridged to create continuous membership periods. Contributed days began after a 180 day baseline period, and ended for that member at the first incident diagnosis of any clinical concept, the last day of enrollment, or the end of the study period (December 31, 2003), whichever came first. Figure 1 illustrates how contributed time was parsed. The baseline period was used as to identify prior diagnoses; no exclusions were applied during baseline.
Table 1. A small subset of the Multi-Level Clinical Classification Tree with International Classification of Diseases, Ninth Revision (ICD-9) codes associated with a specific level.
Table 1. A small subset of the Multi-Level Clinical Classification Tree with International Classification of Diseases, Ninth Revision (ICD-9) codes associated with a specific level.
07Diseases of the Circulatory System
07.01.02  Hypertension with Complications and Secondary Hypertension      Hypertensive Heart and/or Renal Disease (402.00–404.93)      Other Hypertensive Complications (405.01–405.99,437.2)
07.02Diseases of The Heart
07.02.01  Heart Valve Disorders      Nonrheumatic Mitral Valve Disorders (424.0)      Nonrheumatic Aortic Valve Disorders (424.1)      Other Heart Valve Disorders (424.2, 424.3, 785.2, 785.3)
07.02.02  Peri; Endo; and Myocarditis; Cardiomyopathy (Except that Caused by TB or STD)      Cardiomyopathy (425.0–425.9)
07.02.03  Acute Myocardial Infarction (410.0–410.92)
07.02.04  Coronary Atherosclerosis and Other Heart Disease      Angina Pectoris (413.0–413.9)      Unstable Angina (Intermediate Coronary Syndrome) (411.1)      Other Acute and Subacute Forms of Ischemic Heart Disease (411.0, 411.8–411.89)      Coronary Atherosclerosis (414.05)      Other Forms of Chronic Heart Disease (414.8, 414.9)      Other (414.06)
Figure 1. Contributed person time: member timelines.
Figure 1. Contributed person time: member timelines.
Pharmaceutics 05 00179 g001

2.6. Drug Exposure

Contributed days were either exposed or unexposed person time. Treatment episodes (i.e., exposed person time) began the day after a drug dispensing and continued until the end of exposure based on days supplied. Consecutive dispensings were combined, and exposure gaps of six days or less were bridged to create continuous episodes. Unexposed person time was defined as all contributed days without exposure. For each product we calculated total exposed and unexposed person time.

2.7. Outcomes

We defined an incident outcome as the first observed diagnosis during contributed time that was not observed during baseline. Only the first incident outcome observed was counted and designated as exposed or unexposed; this restriction is necessary for the TreeScan analysis that adjusts for multiple testing and was applied here to enable comparison across methods.

2.8. Calculation of Observed and Expected Counts

Exposed outcomes are the number of incident outcomes observed during exposed days. The unadjusted expected count is the number of exposed days times the rate of incident outcomes during unexposed days, calculated as the number of unexposed outcomes divided by the number of unexposed days. Using indirect standardization, we adjusted expected counts for age, sex, and health plan.
Following the distributed data model approach [29,30,31], each site executed analytic code provided by the study coordinating center. Analytic program output contained counts of exposed and unexposed days and outcomes by age (5-year strata) and sex; counts were transferred to the coordinating center for aggregation and analysis.

2.9. Gamma Poisson Shrinker

The GPS was proposed by DuMouchel [37,38] as a signal detection tool for large frequency tables with both observed (O) and expected (E) counts for each drug-outcome pair. It assumes the observed count of any drug-outcome pair follows the Poisson distribution. For spontaneous reports, there are no drug exposure denominator data, so the expected counts are calculated under the null assumption that each drug has the same proportion of diagnosis codes. That is, the expected counts are internally derived assuming the independence of drug and event reporting, and calculated as the product of two marginal frequencies of the drug-outcome pair and the total count of all observed events. For example, if seizures comprise 1% of all the diagnosis codes, over all drugs, and itraconazole has a total of 800 diagnoses, then the expected number of seizures is eight for itraconazole.
Unlike spontaneous reporting databases, population-based event monitoring using health plan data can calculate observed and expected counts based on observed exposure information and diagnoses observed during exposed and unexposed time. GPS can be directly adapted to such settings with the internally derived expected counts replaced by the expected counts constructed using the denominators.
Details of the GPS algorithm have been extensively described [37,38,39,40]. Briefly, for each drug-outcome pair, the primary parameter of interest was the risk ratio. Rather than using the observed over expected (O/E), GPS uses the empirical Bayesian geometric mean (EBGM) posterior distribution of the risk ratio and the surrounding confidence interval for each drug-outcome pair to identify statistical signals of excess risk. To prevent spurious false positives due to implausibly high risk ratios, GPS implements a Bayesian framework that “shrinks” O/E estimates towards a value which is close to the average O/E values for all drug-event pairs at each level of granularity. For these data, that average is about 1.5. GPS accomplishes this by use of an empirical Bayesian framework where the values of all O/E estimates are modeled as a mixture distribution. This so-called “prior distribution” is then combined with data on a specific drug–outcome pair to give a score: the EBGM. Further work would be needed to determine whether shrinkage towards an average value far from one is justified or represents an artificial attribute that might adversely impact performance of the GPS approach. We evaluated each level of the diagnosis tree separately.
We used two signaling thresholds for GPS. The first is the lower bound of the 95% posterior probability interval of 1.5 or more (medium threshold). Since the average O/E for our population was close to 1.5, this threshold mimics an excess risk but is not adjusted for multiple testing. To informally adjust for multiple testing when applying data mining approaches to spontaneous reporting, the U.S. FDA uses the lower bound of the 90% posterior probability interval of EBGM of greater or equal to two as the signal threshold for their spontaneous reporting system [41]. We used this threshold as our most stringent signaling criteria.

2.10. Comparison to Tree-Based Scan Statistic

TreeScan is a signal detection method that simultaneously looks for excess risk in any of a large number of individual cells in a database and in groups of closely related cells, formally adjusting the p-values for the multiple testing inherent in the large number of overlapping diagnosis groups evaluated [6,42,43]. The paper by Kulldorff et al. (2012) details the TreeScan approach for drug safety surveillance [42]. In brief, a hierarchical classification tree is first constructed for the outcomes where related diagnoses are close to each other on the tree. Different cuts on the tree are then made, and it is evaluated whether the group of diagnoses on that branch of the tree has an excess risk of occurring among the drug users. In this way, the method evaluates both very specific outcome definitions such as Paralytic Ileus (a single leaf on the tree) as well as large groups of related outcomes such as Diseases of the Digestive System (one of the largest branches on the tree). The method formally adjusts for the multiple testing inherent in the hundreds or thousands of different cuts evaluated.
For the comparison between GPS and TreeScan we used identical input datasets of age and sex stratified O and E counts for each MLCCS node separately for each drug. We conducted a post-hoc comparison of the GPS and TreeScan results focusing on differences in the number of signals identified overall and by signaling threshold. For this comparison we define two signaling thresholds for TreeScan: multiple testing adjusted p-values of <0.001 and 0.001 < p < 0.05. This comparison was made to help put the GPS results in context using a different statistical signaling method with the same input datasets. We used the product label, medical literature, and clinician input to informally categorize the signals as known, likely confounded, or previously unknown. Formal signal evaluation was beyond the scope of the project.

3. Results

3.1. Terbinafine and Itraconazole

Across all thresholds and nodes on the classification tree we identified 10 GPS terbinafine signals and four itraconazole signals (Table 2). One of the 10 terbinafine signals met the highest GPS threshold and five met the highest TreeScan threshold. Of the four itraconazole signals, one met the highest GPS threshold and two met the highest TreeScan threshold. The antifungal signals represent known AEs (e.g., liver conditions, allergic reactions, nausea) or likely confounding by indication (e.g., skin and subcutaneous tissue diagnoses) [32,44,45], Over 46,000 exposed days for terbinafine, 415,000 exposed days for itraconazole, and 1.1 billion unexposed days were assessed.

3.2. Pioglitazone and Rosiglitazone

Of the 35 pioglitazone signals identified by either method, 15 met the highest GPS signaling threshold and 27 met the highest TreeScan threshold. For the 22 rosiglitazone signals, six met the highest GPS threshold and 15 met the highest TreeScan threshold (Table 3). Most of pioglitazone signals were in four body systems, including signals for Coronary Atherosclerosis and Other Heart Disease (MLCCS node 07.02.04), Congestive Heart Failure (07.02.11 and, and Peripheral and Visceral Atherosclerosis (07.04.01). Pioglitazone also had high threshold GPS signals for Nephritis, Nephrosis and Renal Sclerosis (10.01.01) and Chronic Renal Failure (10.01.03). Over 1.3 million exposed days for pioglitazone, 637,000 exposed days for rosiglitazone, and 1.1 billion unexposed days were assessed.
The cardiovascular and renal signals are known AEs or likely due to confounding. For example, diabetes patients have a higher risk for renal impairment and, since diabetes medications such as metformin and glyburide are contraindicated for those with renal dysfunction, diabetes patients with renal impairment may have been channeled to pioglitazone and rosiglitazone. Both drugs also signaled strongly for chronic ulcer of the skin, and pioglitazone signaled for eye disorders, another likely example of confounding.
Table 2. Results for terbinafine and itraconazole.
Table 2. Results for terbinafine and itraconazole.
ObsExpO/EEBGMGPS SignalTreeScan p-valueObsExpO/EEBGMGPS SignalTreeScan p-value
05Mental Disorders00.60.01.3 . .
06Diseases Of The Nervous System And Sense Organs3722.71.61.6 0.28115.22.11.7 0.54
07Diseases Of The Circulatory System5144.41.11.2 .2110.22.11.7 0.13
07.01  Hypertension12.
07.02  Diseases Of The Heart2421.
07.02.01    Heart Valve Disorders12.
07.02.03    Acute Myocardial Infarction00.
07.02.04    Coronary Atherosclerosis And Other Heart Disease33.
07.02.07    Other And Ill-Defined Heart Disease31.
07.02.08    Conduction Disorders10.
07.02.09    Cardiac Dysrhythmias1610.
07.03  Cerebrovascular Disease42.
07.04  Diseases Of Arteries; Arterioles; And Capillaries1713.
07.05  Diseases Of Veins And Lymphatics54.31.21.4 . 0.47
09Diseases Of The Digestive System6337.21.71.6 0.007158.21.81.6 0.63
09.03  Diseases Of Mouth; Excluding Dental53.
09.04  Upper Gastrointestinal Disorders87.
09.06  Lower Gastrointestinal Disorders10.
09.07  Biliary Tract Disease21.
09.08  Liver Disease143.14.53.5**0.0000510.71.41.5.
09.08.02      Other Liver Diseases143.14.53.37**0.0000510.71.41.5.      Other And Unspecified Liver Disorders142.85.14.1***0.0000210.61.61.5.
09.09  Pancreatic Disorders (Not Diabetes)
09.09.03    Other Pancreatic Disorders20.
09.10  Gastrointestinal Hemorrhage126.
09.12  Other Gastrointestinal Disorders1914.41.31.4 . .
10Diseases Of The Genitourinary System2923.51.21.3 . .
11Complications Of Pregnancy; Childbirth; And The Puerperium00.60.01.2 . .
12Diseases Of The Skin And Subcutaneous Tissue12551.62.42.2**0.000013111.22.82.1**0.00001
12.01  Skin And Subcutaneous Tissue Infections43.
12.02  Other Inflammatory Condition Of Skin2510.
12.03  Chronic Ulcer Of Skin10.
12.04  Other Skin Disorders9537.02.62.3**0.00001198.12.31.8 0.05
13Diseases Of The Musculoskeletal System And Connective Tissue6043.31.41.4 0.59159.11.71.6 0.84
13.01  Infective Arthritis And Osteomyelitis (Except That Caused By TB Or STD10.***0.00001
13.08  Other Connective Tissue Disease5942.71.41.4 0.63118.91.21.4 .
16Injury And Poisoning20.82.61.6 . .
17Symptoms; Signs; And Ill-Defined Conditions And Factors Influencing Health6238.11.61.6 0.02158.51.81.6 0.75
17.01  Symptoms; Signs; And Ill-Defined Conditions6238.
17.01.01    Syncope32.
17.01.06    Nausea And Vomiting103.
17.01.07    Abdominal Pain2118.
17.01.08    Malaise And Fatigue32.
17.01.09    Allergic Reactions2510.82.32.0 0.0152.22.31.7 .
MLCCS = Multi-level Clinical Classifications System; Obs = Observed; Exp = Expected; O/E = Observed/Expected; TreeScan p = multiple testing adjusted p-values, p > 0.90 is indicated with ‘.’; EBGM: empirical Bayesian geometric mean; *** GPS Signal at lower 90% CI bound ≥2; ** Signal at lower 95% CI bound >1.5; Table includes (1) all major disease category headings; (2) any disease categories that signaled for GPS or with a p-value ≤ 0.10; (3) any parents/grandparent of categories that signaled, and (4) any sibling of a disease category that signaled as long as there were observed events. We excluded any categories that were exactly the same as the parent.
Table 3. Results for pioglitazone and rosiglitazone.
Table 3. Results for pioglitazone and rosiglitazone.
ObsExpO/EEBGMGPS SignalTreeScan p-valueObsExpO/EEBGMGPS SignalTreeScan p-value
05Mental Disorders41.72.41.6 . .
06Diseases Of The Nervous System And Sense Organs19790.72.22.1**0.000017545.11.71.6 0.003
06.03  Paralysis21.
06.04  Epilepsy; Convulsions34.
06.05  Headache; Including Migraine411.
06.06  Coma; Stupor; And Brain Damage31.
06.07  Eye Disorders18572.92.52.4***0.000016735.
06.07.01    Cataract12351.42.42.3**0.000014724.
06.07.03    Glaucoma6221.42.92.6***0.000012011.71.71.6 0.80
07Diseases Of The Circulatory System378177.72.12.1**0.0000118694.82.01.9**0.00001
07.01  Hypertension3416.
07.01.02    Hypertension With Complications And Secondary Hypertension3416.      Hypertensive Heart And/Or Renal Disease3416.
07.02  Diseases Of The Heart19086.62.22.1**0.0000111651.92.22.1**0.00001
07.02.01    Heart Valve Disorders1510.
07.02.02    Peri-; Endo-; And Myocarditis; Cardiomyopathy (Except That Caused62.
07.02.03    Acute Myocardial Infarction124.
07.02.04    Coronary Atherosclerosis And Other Heart Disease5118.22.82.4**0.000012410.      Coronary Atherosclerosis And Other Heart Disease113.      Angina Pectoris166.      Unstable Angina (Intermediate Coronary Syndrome)      Other Acute And Subacute Forms Of Ischemic Heart Disease41.      Other Forms Of Chronic Heart Disease155.
07.02.05    Nonspecific Chest Pain21.
07.02.06    Pulmonary Heart Disease11.
07.02.07    Other And Ill-Defined Heart Disease55.
07.02.08    Conduction Disorders22.
07.02.09    Cardiac Dysrhythmias4833.
07.02.10    Cardiac Arrest And Ventricular Fibrillation40.
07.02.11    Congestive Heart Failure; Nonhypertensive447.36.05.9***0.00001203.26.36.0***0.00001      Congestive Heart Failure; Nonhypertensive40.154.614.1***0.0000210.026.42.1.      Congestive Heart Failure356.35.55.4***0.00001162.75.85.3***0.00002      Heart Failure50.
07.03  Cerebrovascular Disease4015.12.62.2**0.00001137.11.81.6.
07.03.01    Acute Cerebrovascular Disease145.
07.03.02    Occlusion Or Stenosis Of Precerebral Arteries72.**0.00472
07.03.03    Other And Ill-Defined Cerebrovascular Disease31.
07.03.04    Transient Cerebral Ischemia143.
07.03.05    Late Effects Of Cerebrovascular Disease20.
07.04  Diseases Of Arteries; Arterioles; And Capillaries8144.
07.04.01    Peripheral And Visceral Atherosclerosis337.64.44.1***0.00001183.84.74.0***0.00002      Atherosclerosis Of Arteries Of Extremities91.65.53.6**0.00530.74.21.9.      Peripheral Vascular Disease Unspecified214.94.33.5***0.00001112.44.73.1**0.0023      Other Peripheral And Visceral Atherosclerosis31.
07.04.02    Aortic; Peripheral; And Visceral Artery Aneurysms11.
07.04.03    Aortic And Peripheral Arterial Embolism Or Thrombosis50.
07.04.04    Other Circulatory Disease4234.      Hypotension81.      Other And Unspecified Circulatory Disease3432.
07.05  Diseases Of Veins And Lymphatics3314.
07.05.01    Phlebitis; Thrombophlebitis And Thromboembolism94.
07.05.02    Varicose Veins Of Lower Extremity55.
07.05.04    Other Diseases Of Veins And Lymphatics195.03.82.9**0.0000231.71.71.5 .
09Diseases Of The Digestive System131116.41.11.2 .8062.01.31.3 0.79
10Diseases Of The Genitourinary System18675.02.52.3***0.000019645.42.12.0**0.00001
10.01  Diseases Of The Urinary System16763.72.62.5***0.000018635.52.42.2**0.00001
10.01.01    Nephritis; Nephrosis; Renal Sclerosis281.224.321.0***0.0000160.78.75.7**0.005
10.01.02    Acute And Unspecified Renal Failure31.
10.01.03    Chronic Renal Failure141.69.08.3***0.0000170.710.17.8***0.0001
10.01.04    Urinary Tract Infections61.
10.01.05    Calculus Of Urinary Tract68.
10.01.06    Other Diseases Of Kidney And Ureters306.24.84.7***0.00001143.44.12.9**0.0004      Other And Unspecified Diseases Of Kidney And Ureters305.55.55.3***0.00001143.04.73.6**0.00004
10.01.07    Other Diseases Of Bladder And Urethra01.
10.01.08    Genitourinary Symptoms And Ill-Defined Conditions8042.**0.00002      Hematuria1311.      Retention Of Urine32.      Other And Unspecified Genitourinary Symptoms6428.12.32.1**0.000013715.92.32.1**0.00007
10.03  Diseases Of Female Genital Organs1911.21.71.6 0.88109.91.01.3 .
11Complications Of Pregnancy; Childbirth; And The Puerperium00.50.01.3 . .
12Diseases Of The Skin And Subcutaneous Tissue205166.61.21.3 0.237870.51.11.2 .
12.01  Skin And Subcutaneous Tissue Infections177.
12.02  Other Inflammatory Condition Of Skin4133.
12.03  Chronic Ulcer Of Skin131.111.810.7***0.0000190.421.216.4***0.00001
12.03.02    Chronic Ulcer Of Leg Or Foot131.111.810.4***0.0000190.421.215.1***0.00001
12.04  Other Skin Disorders134123.81.11.1 .4552.70.91.0 .
13Diseases Of The Musculoskeletal System And Connective Tissue187131.81.41.4 0.000088457.71.51.5 0.07
13.01  Infective Arthritis And Osteomyelitis (Except That Caused By TB Or STD21.
13.07  Systemic Lupus Erythematosus And Connective Tissue Disorders11.
13.08  Other Connective Tissue Disease184129.61.41.4 0.00018156.41.41.4 0.13
16Injury And Poisoning21.81.11.4 . .
17Symptoms; Signs; And Ill-Defined Conditions And Factors Influencing Health128106.81.21.2 .5750.41.11.2 .
MLCCS = Multi-level Clinical Classifications System; Obs = Observed; Exp = Expected; O/E = Observed/Expected; TreeScan p = multiple testing adjusted p-values, p > 0.90 is indicated with ‘.’; EBGM: empirical Bayesian geometric mean; *** GPS Signal at lower 90% CI bound ≥2; **Signal at lower 95% CI bound >1.5; Table includes (1) all major disease category headings; (2) any disease categories that signaled for GPS or with a p-value ≤ 0.10; (3) any parents/grandparent of categories that signaled, and (4) any sibling of a disease category that signaled as long as there were observed events. We excluded any categories that were exactly the same as the parent.

3.3. Overall Signaling Comparison

Table 4 presents all 71 signals. TreeScan identified 71 signals, 49 at the highest threshold (p ≤ 0.001) and 22 at medium threshold (0.001< p < 0.05). GPS identified 48 signals; all high threshold GPS signals were also high threshold TreeScan signals, and 84% (21/25) of the moderate threshold GPS signals were high TreeScan signals. There were five high threshold TreeScan signals that did not signal for GPS, all had low O/E ratios (1.4 to 1.9) and large observed counts (67 to 187).
Table 4. Comparison of all GPS and TreeScan signals across the four drugs studied by threshold.
Table 4. Comparison of all GPS and TreeScan signals across the four drugs studied by threshold.
Tree-Based Scan Statistic (TreeScan) Signal Thresholds
GPS Signal Thresholdsp ≤ 0.0010.001< p ≤ 0.05Total (%)
High23023 (32)
Medium21425 (35)
No signal51823 (32)
GPS Thresholds
Medium: lower 95% CI bound ≥1.5
High: lower 90% CI bound >2.0
Thresholds are mutually exclusive
Note: Data are no. (% of all signals). GPS: Gamma Poisson Shrinker.

4. Discussion

Electronic healthcare databases hold promise for pharmacovigilance because they address common shortcomings inherent to spontaneous reporting systems, offer large sample sizes and the potential to study subgroups, and include longitudinal medical information for a defined population. This is the first study to apply the GPS to population-based observational data in a multi-site environment for assessment of non-prespecified outcomes. Prior studies have implemented GPS in similar environments but have focused on pre-specified drug-event pairs.
We identified 71 signals across four drug products. Of the 48 GPS signals identified, 23 (48%) signaled at the highest threshold. We did not formally evaluate each signal for clinical plausibility or prior knowledge of association, or prioritize them for refinement. All GPS signals were either known associations or could be reasonably attributed to confounding. We counted as unique every signal at each level of the hierarchical tree, an approach that overestimates the number of signals that would require refinement under a real-world implementation. For example, we counted signals for Diseases of the Nervous System and Sense Organs (06), Eye Disorders (06.07), Cataracts (06.07.01), and Glaucoma (06.07.03) as four distinct signals, although the higher-level signals (06 and 06.07) were almost entirely made up of the two lower-level signals and would not require four distinct signal refinement activities.
The GPS and TreeScan results were similar with respect to the clinical areas that signaled, although findings varied by signaling threshold. In the few cases that GPS did not signal and TreeScan signaled at highest threshold, observed counts were high and O/E was low. This is expected behavior based on the nature of the two methods.
We note that using observational data for signal detection requires complex and often subjective analytic specifications, often unstated, that can directly affect the interpretation of the findings. Our implementation applied common epidemiologic approaches such as allowing individuals to contribute exposed and unexposed person time, and most importantly, involved identification of non-specified incident outcomes. The pilot was conducted using a fully distributed approach that did not require sharing of person-level protected health data but allowed identical definitions and analyses to be performed at each data partner.
Implementation required specifications such as the baseline period, the allowable enrollment gap, the definition of contributed, exposed, and unexposed time, the creation of treatment episodes, allowable treatment gaps, right censoring decisions, and incident outcome definitions. A longer baseline period could reduce confounding by indication by identifying more patients with prevalent comorbid conditions, but would reduce overall cohort size due to lack of complete baseline. The incident outcome also can be defined several ways taking into account baseline period, care setting (inpatient versus outpatient), and diagnosis coding hierarchy. By excluding diagnoses unlikely to be associated with an acute outcome we eliminated the possibility of identifying the excluded diagnoses as outcomes. Our exclusion of injury and poisoning codes is debatable; those codes could be valuable in identifying those risks, but there is some uncertainty about how well those codes reflect adverse events versus misuse of safe medications (e.g., overdose). Regardless, they could be added in future implementations, especially in combination with loosening our simplifying restriction that allowed only one incident event per person. Finally, we compared exposed to unexposed time, controlling for age group, sex, and health plan. Others options for confounding adjustment include using an exposed comparison cohort, narrower age stratifications, and matching or disease severity stratification.
Further, we note that we did not identify a set of expected signals that we hoped to find (e.g., expected adverse drug events) or avoid (e.g., confounded signals). Rather, since others have identified GPS as a potentially valuable tool for signal detection, our goal was to investigate the feasibility of implementing GPS and TreeScan in a real-world large-scale data mining application using longitudinal electronic health data without limiting the analysis to pre-specified relationships. We reviewed the product label, medical literature, and consulted clinical experts (co-authors and others) to informally assess whether identified signals were known or reasonably expected due to confounding. In our view there is no clear and well defined list of all known adverse drug events. Most drug labels have extensive adverse event list, but there is no conclusive evidence that all these “adverse events” are caused by the drug. For black box warnings there is usually strong evidence, but most known adverse events do not generate a black box warning. Developing a comprehensive list of all possible adverse events was beyond the scope of the paper and inconsistent with our primary aim to assess implementation in a real world setting.
There is no “correct” decision regarding signaling thresholds, only generally understood trade-offs between identifying more or fewer signals and the resulting changes in numbers of signals needing refinement. The ability to interpret the signal detection findings within the larger pharmacovigilance framework, the ability of the findings to inform decision-making, and effort needed to evaluate them are important factors in whether or not a surveillance approach is viable. For instance, a viable approach will generate signals with enough informational value that they can be quickly adjudicated as likely due to confounding, known, expected, or otherwise uninteresting versus those that require further evaluation and refinement.
We have shown that the GPS method can be successfully applied to population-based health plan data and that it performs adequately, with the ability to detect known adverse events. Although the GPS always shrinks the O/E estimate to some value to reduce variability, it is not always the case that the O/E shrinks towards 1.0 [38]. We observed shrinkage towards approximately 1.5. This property could be seen as less than desirable. It reflects that across all the outcome events under study, there is an excess number of events compared to the expected. Thus, the O/E is shrunk towards some average O/E taken over all outcomes. The fact that this is more than 1.0 could be because the drug causes a whole range of adverse events, or more likely, that the drug is taken by a generally sicker or more frail population that experiences a whole range of comorbidities. Prior implementations of GPS in longitudinal data did not report this finding or any metrics regarding the prior distributions. This finding emphasizes the need for research with empirical Bayesian approaches to report details of the prior distribution so that it is clear to which value shrinkage occurs, and in general the importance of transparency of reported findings, around for example, signaling thresholds and the need to consider alternative implementation approaches such as the use of zero-inflated Poisson to account for the many empty cells and underlying variability. More research regarding the appropriate implementation strategy for GPS using longitudinal data and signaling thresholds and strategies for multi-level testing are also needed, including the potential to adjust thresholds and approaches based on the specific surveillance target(s) and perhaps observed prior distributions.
Compared with TreeScan, there are both strengths and weaknesses to GPS, and it may be ideal to employ both methods simultaneously, using the combined results to better strengthen, refute, and understand signals. A strength of the GPS method is the Bayesian probability intervals; shrinkage of the point estimates towards the population average of O/E is a possible strength, although care must be taken to insure “shrinkage” towards values over unity does not introduce signals. We suggest future implementations carefully assess the GPS parameter sets to understand the distribution of O/E in the underlying population. A major strength of the TreeScan is the formal adjustment for multiple testing and the ability to analyze different levels of disease granularity in a single combined analysis.
Signal detection is one step in the continuum of medical product safety surveillance. Approaches that generate statistical signals difficult to refine are of little practical use [46]—potentially creating an unfortunate situation in which new signal detection methods generate more heat than light. Therefore, it is critical that a robust medical product safety surveillance system have efficient mechanisms—such as those proposed by diverse stakeholders including the governmental agencies, academia and the pharmaceutical industry to quickly prioritize, refine and evaluate and act on signals. Examples of novel approaches to surveillance and signal management include the work of the US FDA Sentinel System [31,47,48] the WHO Collaborating Centre for International Drug Monitoring [49], the Asian Pharmacoepidemiology Network [50], and the European Network of Centres for Pharmacoepidemiology and Pharmacovigilance [51], and are being tested in multiple major international initiatives [52,53,54].


This work was funded by Pfizer Inc. (JSB, FZ, ID, TRA) and the Agency for Healthcare Research and Quality (JSB, ID, MK, KAC, RLD, SEA, DB, MJG, LH, PP, MAR, DR, DS, RP), through a grant to the HMO Research Network Center for Education and Research on Therapeutics (CERT), grant number U18 HS 010391. KP, AB, and RR are employees of Pfizer Inc.
We are indebted to the statistical programmers at each of the sites for their work in extracting the study data and testing the programming algorithms, to Kimberly Lane and the HMO Research Network CERT Data Coordinating Center for their help in overseeing the study.

Conflicts of Interest

The authors declare no conflicts of interest.


  1. Bate, A.; Evans, S.J. Quantitative signal detection using spontaneous ADR reporting. Pharmacoepidemiol. Drug Saf. 2009, 18, 427–436. [Google Scholar] [CrossRef]
  2. Curtis, J.R.; Cheng, H.; Delzell, E.; Fram, D.; Kilgore, M.; Saag, K.; Yun, H.; Dumouchel, W. Adaptation of Bayesian data mining algorithms to longitudinal claims data: Coxib safety as an example. Med. Care 2008, 46, 969–975. [Google Scholar] [CrossRef]
  3. Schuemie, M.J. Methods for drug safety signal detection in longitudinal observational databases: LGPS and LEOPARD. Pharmacoepidemiol. Drug Saf. 2011, 20, 292–299. [Google Scholar] [CrossRef]
  4. Norén, G.N.; Bate, A.; Hopstadius, J.; Star, K.; Edwards, I.R. Temporal Pattern Discovery for Treands and Transient Effects: Its Application to Patient Records, Proceedings of The 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Las Vegas, NV, USA, August 24–27, 2008; pp. 963–971.
  5. Brown, J.S.; Petronis, K.; Bate, A.; Zhang, F.; Dashevsky, I.; Kulldorff, M.; Avery, T.A.; Davis, R.L.; Andrade, S.E.; Dublin, S.; et al. Comparing Two Methods for Detecting Adverse Event Signals in Observational Data: Empirical Bayes Gamma Poisson Shrinker and Tree-Based Scan Statistic, Proceedings of The 27th International Conference on Pharmacoepidemiology and Therapeutic Risk Management, The International Society for Pharmacoepidemiology Chicago, Chicago, IL, USA, August 14–17, 2011; p. Abstract 575.
  6. Kulldorff, M.; Dashevsky, I.; Avery, T.A.; Chan, K.A.; Davis, R.L.; Graham, D.; Platt, R.; Andrade, S.E.; Boudreau, D.; Gunter, M.J.; et al. Drug Safety Data Mining with a Tree-Based Scan Statistic, Proceedings of 26th International Conference on Pharmacoepidemiology and Therapeutic Risk Management, Brighton, England, UK, 2010.
  7. Park, M.Y.; Yoon, D.; Lee, K.; Kang, S.Y.; Park, I.; Lee, S.H.; Kim, W.; Kam, H.J.; Lee, Y.H.; Kim, J.H.; et al. A novel algorithm for detection of adverse drug reaction signals using a hospital electronic medical record database. Pharmacoepidemiol. Drug Saf. 2011, 20, 598–607. [Google Scholar] [CrossRef]
  8. Jin, H.W.; Chen, J.; He, H.; Williams, G.J.; Kelman, C.; O’Keefe, C.M. Mining unexpected temporal associations: Applications in detecting adverse drug reactions. IEEE Trans. Inf. Technol. Biomed. 2008, 12, 488–500. [Google Scholar] [CrossRef]
  9. Walker, A.M. Signal detection for vaccine side effects that have not been specified in advance. Pharmacoepidemiol. Drug Saf. 2010, 19, 311–317. [Google Scholar] [CrossRef]
  10. Harpaz, R.; DuMouchel, W.; Shah, N.H.; Madigan, D.; Ryan, P.; Friedman, C. Novel data-mining methodologies for adverse drug event discovery and analysis. Clin. Pharmacol. Ther. 2012, 91, 1010–1021. [Google Scholar] [CrossRef]
  11. Zorych, I.; Madigan, D.; Ryan, P.; Bate, A. Disproportionality methods for pharmacovigilance in longitudinal observational databases. Stat. Methods Med. Res. 2011, 22, 39–56. [Google Scholar] [CrossRef]
  12. Norén, G.N.; Hopstadius, J.; Bate, A.; Star, K.; Edwards, I.R. Temporal pattern discovery in longitudinal electronic patient records. Data Min. Knowl. Discov. 2010, 20, 361–387. [Google Scholar] [CrossRef]
  13. Schuemie, M.J.; Coloma, P.M.; Straatman, H.; Herings, R.M.; Trifiro, G.; Matthews, J.N.; Prieto-Merino, D.; Molokhia, M.; Pedersen, L.; Gini, R.; et al. Using electronic health care records for drug safety signal detection: a comparative evaluation of statistical methods. Med. Care 2012, 50, 890–897. [Google Scholar] [CrossRef]
  14. Norén, G.N.; Hopstadius, J.; Bate, A.; Edwards, I.R. Safety surveillance of longitudinal databases: Results on real-world data. Pharmacoepidemiol. Drug Saf. 2012, 21, 673–675. [Google Scholar] [CrossRef]
  15. Noren, G.N.; Hopstadius, J.; Bate, A.; Edwards, I.R. Safety surveillance of longitudinal databases: Methodological considerations. Pharmacoepidemiol. Drug Saf. 2011, 20, 714–717. [Google Scholar] [CrossRef]
  16. Schuemie, M.J. Safety surveillance of longitudinal databases: further methodological considerations. Pharmacoepidemiol. Drug Saf. 2012, 21, 670–672. [Google Scholar] [CrossRef]
  17. Bate, A.; Brown, E.G.; Goldman, S.A.; Hauben, M. Terminological challenges in safety surveillance. Drug Saf. 2012, 35, 79–84. [Google Scholar] [CrossRef]
  18. Ryan, P.B.; Madigan, D.; Stang, P.E.; Overhage, J.M.; Racoosin, J.A.; Hartzema, A.G. Empirical assessment of methods for risk identification in healthcare data: Results from the experiments of the Observational Medical Outcomes Partnership. Stat. Med. 2012, 31, 4401–4415. [Google Scholar] [CrossRef]
  19. Vaccine Safety Monitoring—Adverse Events. Available online: (accessed on 15 August 2012).
  20. Yih, W.K.; Kulldorff, M.; Fireman, B.H.; Shui, I.M.; Lewis, E.M.; Klein, N.P.; Baggs, J.; Weintraub, E.S.; Belongia, E.A.; Naleway, A.; et al. Active surveillance for adverse events: The experience of the Vaccine Safety Datalink project. Pediatrics 2011, 127, S54–S64. [Google Scholar] [CrossRef]
  21. Platt, R.; Davis, R.; Finkelstein, J.; Go, A.S.; Gurwitz, J.H.; Roblin, D.; Soumerai, S.; Ross-Degnan, D.; Andrade, S.; Goodman, M.J.; et al. Multicenter epidemiologic and health services research on therapeutics in the HMO Research Network Center for Education and Research on Therapeutics. Pharmacoepidemiol. Drug Saf. 2001, 10, 373–377. [Google Scholar] [CrossRef]
  22. Platt, R.; Andrade, S.E.; Davis, R.L.; Destefano, F.; Finkelstein, J.A.; Goodman, M.J.; Gurwitz, J.Y.; Go, A.S.; Martinson, B.C.; Raebel, M.A.; et al. Pharmacovigilance in the HMO Research Network. In Pharmacovigilance; Mann, R.D., Andrews, E.B., Eds.; Wiley: New York, NY, USA, 2002; pp. 392–398. [Google Scholar]
  23. Raebel, M.A.; Lyons, E.E.; Andrade, S.E.; Chan, K.A.; Chester, E.A.; Davis, R.L.; Ellis, J.L.; Feldstein, A.; Gunter, M.J.; Lafata, J.E.; et al. Laboratory monitoring of drugs at initiation of therapy in ambulatory care. J. Gen. Intern. Med. 2005, 20, 1120–1126. [Google Scholar] [CrossRef]
  24. Raebel, M.A.; McClure, D.L.; Simon, S.R.; Chan, K.A.; Feldstein, A.; Andrade, S.E.; Lafata, J.E.; Roblin, D.; Davis, R.L.; Gunter, M.J.; et al. Laboratory monitoring of potassium and creatinine in ambulatory patients receiving angiotensin converting enzyme inhibitors and angiotensin receptor blockers. Pharmacoepidemiol. Drug Saf. 2007, 16, 55–64. [Google Scholar] [CrossRef]
  25. Simon, S.R.; Andrade, S.E.; Ellis, J.L.; Nelson, W.W.; Gurwitz, J.H.; Lafata, J.E.; Davis, R.L.; Feldstein, A.; Raebel, M.A. Baseline laboratory monitoring of cardiovascular medications in elderly health maintenance organization enrollees. J. Am. Geriatr. Soc. 2005, 53, 2165–2169. [Google Scholar] [CrossRef]
  26. Simon, S.R.; Chan, K.A.; Soumerai, S.B.; Wagner, A.K.; Andrade, S.E.; Feldstein, A.C.; Lafata, J.E.; Davis, R.L.; Gurwitz, J.H. Potentially inappropriate medication use by elderly persons in U.S. Health Maintenance Organizations, 2000–2001. J. Am. Geriatr. Soc. 2005, 53, 227–232. [Google Scholar]
  27. Wagner, A.K.; Chan, K.A.; Dashevsky, I.; Raebel, M.A.; Andrade, S.E.; Lafata, J.E.; Davis, R.L.; Gurwitz, J.H.; Soumerai, S.B.; Platt, R. FDA drug prescribing warnings: is the black box half empty or half full? Pharmacoepidemiol. Drug Saf. 2006, 15, 369–386. [Google Scholar] [CrossRef]
  28. Chan, J.; Hui, R.L.; Levin, E. Differential association between statin exposure and elevated levels of creatine kinase. Ann. Pharmacother. 2005, 39, 1611–1616. [Google Scholar] [CrossRef]
  29. Velentgas, P.; Bohn, R.L.; Brown, J.S.; Chan, K.A.; Gladowski, P.; Holick, C.N.; Kramer, J.M.; Nakasato, C.; Spettell, C.M.; Walker, A.M.; et al. A distributed research network model for post-marketing safety studies: The Meningococcal Vaccine Study. Pharmacoepidemiol. Drug Saf. 2008, 17, 1226–1234. [Google Scholar] [CrossRef]
  30. Brown, J.; Moore, K.; Braun, M.; Ziyadeh, N.; Chan, K.; Lee, G.; Kulldorff, M.; Walker, A.; Platt, R. Active influenza vaccine safety surveillance: Potential within a healthcare claims environment. Med. Care 2009, 47, 1251–1257. [Google Scholar] [CrossRef]
  31. Behrman, R.E.; Benner, J.S.; Brown, J.S.; McClellan, M.; Woodcock, J.; Platt, R. Developing the Sentinel System—A national resource for evidence development. N. Engl. J. Med. 2011, 364, 498–499. [Google Scholar] [CrossRef]
  32. Perveze, Z.; Johnson, M.W.; Rubin, R.A.; Sellers, M.; Zayas, C.; Jones, J.L.; Cross, R.; Thomas, K.; Butler, B.; Shrestha, R. Terbinafine-induced hepatic failure requiring liver transplantation. Liver Transpl. 2007, 13, 162–164. [Google Scholar] [CrossRef]
  33. Lou, H.Y.; Fang, C.L.; Fang, S.U.; Tiong, C.; Cheng, Y.C.; Chang, C.C. Hepatic failure related to itraconazole use successfully treated by corticosteroids. Hepat. Mon. 2011, 11, 843–846. [Google Scholar]
  34. Nikkels, A.F.; Nikkels-Tassoudji, N.; Pierard, G.E. Oral antifungal-exacerbated inflammatory flare-up reactions of dermatomycosis: Case reports and review of the literature. Am. J. Clin. Dermatol. 2006, 7, 327–331. [Google Scholar] [CrossRef]
  35. Cançado, G.G.; Fujiwara, R.T.; Freitas, P.A.; Correa-Oliveira, R.; Bethony, J.M. Acute generalized exanthematous pustulosis induced by itraconazole: an immunological approach. Clin. Exp. Dermatol. 2009, 34, e709–e711. [Google Scholar] [CrossRef]
  36. Elixhauser, A.; Steiner, C.; Palmer, L. Clinical Classifications Software (CCS), 2009. Agency for Healthcare Research and Quality. Available online: (accessed on 15 August 2012).
  37. DuMouchel, W. Bayesian data mining in large frequency tables, with an application to the FDA spontaneous reporting system. Am. Stat. 1999, 53, 177–190. [Google Scholar]
  38. Fram, D.; Almenoff, J.S.; Dumouchel, W. Empirical Bayesian Data Mining for Discovering Patterns in Post-Marketing Drug Safety. In Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington DC, USA, August 24−27, 2003; pp. 359–368.
  39. Banks, D.; Woo, E.J.; Burwen, D.R.; Perucci, P.; Braun, M.M.; Ball, R. Comparing data mining methods on the VAERS database. Pharmacoepidemiol. Drug Saf. 2005, 14, 601–609. [Google Scholar] [CrossRef]
  40. Almenoff, J.S.; DuMouchel, W.; Kindman, L.A.; Yang, X.; Fram, D. Disproportionality analysis using empirical Bayes data mining: A tool for the evaluation of drug interactions in the post-marketing setting. Pharmacoepidemiol. Drug Saf. 2003, 12, 517–521. [Google Scholar] [CrossRef]
  41. Szarfman, A.; Machado, S.G.; O’Neill, R.T. Use of screening algorithms and computer systems to efficiently signal higher-than-expected combinations of drugs and events in the US FDA’s spontaneous reports database. Drug Saf. 2002, 25, 381–392. [Google Scholar] [CrossRef]
  42. Kulldorff, M.; Dashevsky, I.; Avery, T.; Chan, A.; Davis, R.; Graham, D.; Platt, R.; Andrade, S.; Boudreau, D.; Dublin, S.; et al. Drug Safety Data Mining with a Tree-Based Scan Statistic. Pharmacoepidemiol. Drug Saf. 2013, in press. [Google Scholar]
  43. Kulldorff, M.; Fang, Z.; Walsh, S.J. A tree-based scan statistic for database disease surveillance. Biometrics 2003, 59, 323–331. [Google Scholar] [CrossRef]
  44. Tuccori, M.; Bresci, F.; Guidi, B.; Blandizzi, C.; Del Tacca, M.; Di Paolo, M. Fatal hepatitis after long-term pulse itraconazole treatment for onychomycosis. Ann. Pharmacother. 2008, 42, 1112–1117. [Google Scholar] [CrossRef]
  45. Kohli, R.; Hadley, S. Fungal arthritis and osteomyelitis. Infect. Dis. Clin. North Am. 2005, 19, 831–851. [Google Scholar] [CrossRef]
  46. Avorn, J.; Schneeweiss, S. Managing drug-risk information—What to do with all those new numbers. N. Engl. J. Med. 2009, 361, 647–649. [Google Scholar] [CrossRef]
  47. Robb, M.A.; Racoosin, J.A.; Sherman, R.E.; Gross, T.P.; Ball, R.; Reichman, M.E.; Midthun, K.; Woodcock, J. The US Food and Drug Administration’s Sentinel Initiative: Expanding the horizons of medical product safety. Pharmacoepidemiol. Drug Saf. 2012, 21, 9–11. [Google Scholar] [CrossRef]
  48. Platt, R.; Carnahan, R.M.; Brown, J.S.; Chrischilles, E.; Curtis, L.H.; Hennessy, S.; Nelson, J.C.; Racoosin, J.A.; Robb, M.; Schneeweiss, S.; et al. The U.S. Food and Drug Administration’s Mini-Sentinel program: Status and direction. Pharmacoepidemiol. Drug Saf. 2012, 21, 1–8. [Google Scholar]
  49. Olsson, S. The role of the WHO programme on International Drug Monitoring in coordinating worldwide drug safety efforts. Drug Saf. 1998, 19, 1–10. [Google Scholar] [CrossRef]
  50. Asian Pharmacoepidemiology Network. Available online: (accessed on 21 November 2012).
  51. Blake, K.V.; Devries, C.S.; Arlett, P.; Kurz, X.; Fitt, H. Increasing scientific standards, independence and transparency in post-authorisation studies: The role of the European Network of Centres for Pharmacoepidemiology and Pharmacovigilance. Pharmacoepidemiol. Drug Saf. 2012, 21, 690–696. [Google Scholar] [CrossRef]
  52. Stang, P.E.; Ryan, P.B.; Racoosin, J.A.; Overhage, J.M.; Hartzema, A.G.; Reich, C.; Welebob, E.; Scarnecchia, T.; Woodcock, J. Advancing the science for active surveillance: Rationale and design for the Observational Medical Outcomes Partnership. Ann. Intern. Med. 2010, 153, 600–606. [Google Scholar]
  53. Coloma, P.M.; Schuemie, M.J.; Trifiro, G.; Gini, R.; Herings, R.; Hippisley-Cox, J.; Mazzaglia, G.; Giaquinto, C.; Corrao, G.; Pedersen, L.; et al. Combining electronic healthcare databases in Europe to allow for large-scale drug safety monitoring: the EU-ADR Project. Pharmacoepidemiol. Drug Saf. 2011, 20, 1–11. [Google Scholar] [CrossRef]
  54. Pharmacoepidemiological Research on Outcomes of Therapeutics by a European Consortium IMI-PROTECT. Available online: (accessed on 21 November 2012).

Share and Cite

MDPI and ACS Style

Brown, J.S.; Petronis, K.R.; Bate, A.; Zhang, F.; Dashevsky, I.; Kulldorff, M.; Avery, T.R.; Davis, R.L.; Chan, K.A.; Andrade, S.E.; Boudreau, D.; Gunter, M.J.; Herrinton, L.; Pawloski, P.A.; Raebel, M.A.; Roblin, D.; Smith, D.; Reynolds, R. Drug Adverse Event Detection in Health Plan Data Using the Gamma Poisson Shrinker and Comparison to the Tree-based Scan Statistic. Pharmaceutics 2013, 5, 179-200.

AMA Style

Brown JS, Petronis KR, Bate A, Zhang F, Dashevsky I, Kulldorff M, Avery TR, Davis RL, Chan KA, Andrade SE, Boudreau D, Gunter MJ, Herrinton L, Pawloski PA, Raebel MA, Roblin D, Smith D, Reynolds R. Drug Adverse Event Detection in Health Plan Data Using the Gamma Poisson Shrinker and Comparison to the Tree-based Scan Statistic. Pharmaceutics. 2013; 5(1):179-200.

Chicago/Turabian Style

Brown, Jeffrey S., Kenneth R. Petronis, Andrew Bate, Fang Zhang, Inna Dashevsky, Martin Kulldorff, Taliser R. Avery, Robert L. Davis, K. Arnold Chan, Susan E. Andrade, Denise Boudreau, Margaret J. Gunter, Lisa Herrinton, Pamala A. Pawloski, Marsha A. Raebel, Douglas Roblin, David Smith, and Robert Reynolds. 2013. "Drug Adverse Event Detection in Health Plan Data Using the Gamma Poisson Shrinker and Comparison to the Tree-based Scan Statistic" Pharmaceutics 5, no. 1: 179-200.

Article Metrics

Back to TopTop