Alan M. Hochberg

An Evaluation of Three Signal-Detection Algorithms Using a Highly Inclusive Reference Event Database

AbstractBackground: Pharmacovigilance data-mining algorithms (DMAs) are known to generate significant numbers of false-positive signals of disproportionate reporting (SDRs), using various standards to define the terms ‘true positive’ and ‘false positive’. Objective: To construct a highly inclusive reference event database of reported adverse events for a limited set of drugs, and to utilize that database to evaluate three DMAs for their overall yield of scientifically supported adverse drug effects, with an emphasis on ascertaining false-positive rates as defined by matching to the database, and to assess the overlap among SDRs detected by various DMAs. Methods: A sample of 35 drugs approved by the US FDA between 2000 and 2004 was selected, including three drugs added to cover therapeutic categories not included in the original sample. We compiled a reference event database of adverse event information for these drugs from historical and current US prescribing information, from peer-reviewed literature covering 1999 through March 2006, from regulatory actions announced by the FDA and from adverse event listings in the British National Formulary. Every adverse event mentioned in these sources was entered into the database, even those with minimal evidence for causality. To provide some selectivity regarding causality, each entry was assigned a level of evidence based on the source of the information, using rules developed by the authors. Using the FDA adverse event reporting system data for 2002 through 2005, SDRs were identified for each drug using three DMAs: an urn-model based algorithm, the Gamma Poisson Shrinker (GPS) and proportional reporting ratio (PRR), using previously published signalling thresholds. The absolute number and fraction of SDRs matching the reference event database at each level of evidence was determined for each report source and the data-mining method. Overlap of the SDR lists among the various methods and report sources was tabulated as well. Results: The GPS algorithm had the lowest overall yield of SDRs (763), with the highest fraction of events matching the reference event database (89 SDRs, 11.7%), excluding events described in the prescribing information at the time of drug approval. The urn model yielded more SDRs (1562), with a non-significantly lower fraction matching (175 SDRs, 11.2%). PRR detected still more SDRs (3616), but with a lower fraction matching (296 SDRs, 8.2%). In terms of overlap of SDRs among algorithms, PRR uniquely detected the highest number of SDRs (2231, with 144, or 6.5%, matching), followed by the urn model (212, with 26, or 12.3%, matching) and then GPS (0 SDRs uniquely detected). Conclusions: The three DMAs studied offer significantly different tradeoffs between the number of SDRs detected and the degree to which those SDRs are supported by external evidence. Those differences may reflect choices of detection thresholds as well as features of the algorithms themselves. For all three algorithms, there is a substantial fraction of SDRs for which no external supporting evidence can be found, even when a highly inclusive search for such evidence is conducted.

Using Data Mining to Predict Safety Actions from FDA Adverse Event Reporting System Data

Purpose: To determine the value of data mining in early identification of drug safety signals from spontaneous reporting databases. Methods: A single data mining algorithm was applied to the 2001–2003 public release of Food and Drug Administration Adverse Event Reporting System (AERS) data for all therapeutic new molecular entities (NMEs) approved in 2001. The list of detected signals was compared with the list of safety-related regulatory actions for those drugs through February 2006. Results: For the 21 NMEs, 73 signals of interest were detected by data mining. In 39 cases, that signal preceded regulatory action. The median time from approval to signal detection was 11.5 months, and the median time from signal detection to action was 21 months. There were 33 actions for which no signal was detected and 34 signals with no corresponding regulatory action. Conclusion: Using AERS data 2–3 years following approval, more than half of FDA actions that occurred in the next 2–4 years were predicted by data mining, and more than half of the signals detected by data mining corresponded to an FDA action. An appropriate data mining procedure can yield meaningful safety information, often well in advance of regulatory action.

An experimental investigation of masking in the US FDA adverse event reporting system database.

AbstractBackground: A phenomenon of ‘masking’ or ‘cloaking’ in pharmacovigilance data mining has been described, which can potentially cause signals of disproportionate reporting (SDRs) to be missed, particularly in pharmaceutical company databases. Masking has been predicted theoretically, observed anecdotally or studied to a limited extent in both pharmaceutical company and health authority databases, but no previous publication systematically assesses its occurrence in a large health authority database. Objective: To explore the nature, extent and possible consequences of masking in the US FDA Adverse Event Reporting System (AERS) database by applying various experimental unmasking protocols to a set of drugs and events representing realistic pharmacovigilance analysis conditions. Methods: This study employed AERS data from 2001 through 2005. For a set of 63 Medical Dictionary for Regulatory Activities (MedDRA®) Preferred Terms (PTs), disproportionality analysis was carried out with respect to all drugs included in the AERS database, using a previously described urnmodel-based algorithm. We specifically sought masking in which drug removal induced an increase in the statistical representation of a drug-event combination (DEC) that resulted in the emergence of a new SDR. We performed a series of unmasking experiments selecting drugs for removal using rational statistical decision rules based on the requirement of a reporting ratio (RR) >1, top-ranked statistical unexpectedness (SU) and relatedness as reflected in the WHO Anatomical Therapeutic Chemical level 4 (ATC4) grouping. In order to assess the possible extent of residual masking we performed two supplemental purely empirical analyses on a limited subset of data. This entailed testing every drug and drug group to determine which was most influential in uncovering masked SDRs. We assessed the strength of external evidence for a causal association for a small number of masked SDRs involving a subset of 29 drugs for which level of evidence adjudication was available from a previous study. Results: The original disproportionality analysis identified 8719 SDRs for the 63 PTs. The SU-based unmasking protocols generated variable numbers of masked SDRs ranging from 38 to 156, representing a 0.43–1.8% increase over the number of baseline SDRs. A significant number of baseline SDRs were also lost in the course of our experiments. The trend in the number of gained SDRs per report removed was inversely related to the number of lost SDRs per protocol. Both the number and nature of the reports removed influenced the number of gained SDRs observed. The purely empirical protocols unmasked up to ten times as many SDRs. None of the masked SDRs had strong external evidence supporting a causal association. Most involved associations for which there was no external supporting evidence or were in the original product label. For two masked SDRs, there was external evidence of a possible causal association. Conclusions: We documented masking in the FDA AERS database. Attempts at unmasking SDRs using practically implementable protocols produced only small changes in the output of SDRs in our analysis. This is undoubtedly related to the large size and diversity of the database, but the complex inter-dependencies between drugs and events in authentic spontaneous reporting system (SRS) databases, and the impact of measures of statistical variability that are typically used in real-world disproportionality analysis, may be additional factors that constrain the discovery of masked SDRs and which may also operate in pharmaceutical company databases. Empirical determination of the most influential drugs may uncover significantly more SDRs than protocols based on predetermined statistical selection rules but are impractical except possibly for evaluating specific events. Routine global exercises to elicit masking, especially in large health authority databases are not justified based on results available to date. Exercises to elicit unmasking should be driven by prior knowledge or obvious data imbalances.

Drug-versus-Drug Adverse Event Rate Comparisons : A Pilot Study Based on Data from the US FDA Adverse Event Reporting System

AbstractBackground: A number of published studies compare adverse event rates for drugs on the basis of reports in the US FDA Adverse Event Reporting System (AERS). While the AERS data have the advantage of timely availability and a large capture population, the database is subject to many significant biases, and lacks complete patient information that would allow for correction of those biases. The accuracy of comparative AERS-based data mining has been questioned, but has not been systematically studied. Objective: To determine whether AERS could be used as a data source to accurately compare the adverse event rates for pairs of drugs, using predefined, stringent criteria to dictate whether a given pair of drugs was considered eligible for such a comparison. Methods: The Fisher’s Exact test was utilized to detect differences in adverse event rates between such pairs of drugs. Concordance was determined between statistically significant AERS-based adverse event rate differences, and adverse event rate differences published in the literature from clinical trials and case-control studies. The conditions for validity included (i) data that are free of ‘extreme duplication’ in AERS reports; (ii) drugs used in similar patient populations; (iii) drugs used for similar indications; (iv) drugs used with the same spectrum of concomitant medications; and (v) drugs not widely disparate in time on the market. Results: For 19 drugs studied, a total of 36 evaluable adverse event rate comparisons were identified. Comparisons were classified as favouring ‘drug A’, favouring ‘drug B’ or detecting no difference. Concordance for the resulting 3×3 table (AERS vs literature) gave a kappa statistic of 0.654, indicating moderately good agreement. In only two cases was there absolute discordance, with AERS designating one drug as having a lower rate, while the published study designated the other drug as having a lower rate, with respect to a given adverse event. Conclusions: This pilot study encourages further research regarding the use of spontaneous report databases such as AERS, under stringently defined conditions, to compare adverse event rates for drugs. While not hypothesis proving, such estimates can be used for purposes such as generating hypotheses for controlled studies, and for designing those studies.

The Importance of Reporting Negative Findings in Data Mining

The US Food and Drug Administration (FDA) recently published a warning regarding pancreatitis in association with the use of exenatide, an incretin mimetic used for the treatment of patients with diabetes mellitus. We note that this safety issue is not associated with a signal of disproportionate reporting (SDR) in the FDA Adverse Event Reporting System (AERS) database or the World Health Organization (Uppsala Monitoring Centre) Vigibase for any of four data-mining algorithms we tested (proportional reporting ratio, the multi-item gamma-Poisson shrinker, an urn model and the Bayesian Confidence Propagation Neural Network). Exenatide and acute pancreatitis may thus represent a ‘false-negative’ result for disproportionality-based data-mining methodology generally. We evaluate the possibility that this lack of an SDR is caused by the phenomenon known as ‘masking’ (or ‘cloaking’) and reject this hypothesis. While positive findings are understandably more exciting, we discuss why publishing negative findings, such as in this example, is important for placing the capabilities and limitations of drug safety data mining into proper perspective.

Disease progression modeling from historical clinical databases

This paper considers the problem of modeling disease progression from historical clinical databases, with the ultimate objective of stratifying patients into groups with clearly distinguishable prognoses or suitability for different treatment strategies. To meet this objective, we describe a procedure that first fits clinical variables measured over time to a disease progression model. The resulting parameter estimates are then used as the basis for a stepwise clustering procedure to stratify patients into groups with distinct survival characteristics. As a practical illustration, we apply this procedure to survival prediction, using a liver transplant database from the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK).

Systematic Investigation of Time Windows for Adverse Event Data Mining for Recently Approved Drugs

The optimum timing of drug safety data mining for a new drug is uncertain. The objective of this study was to compare cumulative data mining versus mining with sliding time windows. Adverse Event Reporting System data (2001–2005) were studied for 27 drugs. A literature database was used to evaluate signals of disproportionate reporting (SDRs) from an urn model data‐mining algorithm. Data mining was applied cumulatively and with sliding time windows from 1 to 4 years in width. Time from SDR generation to the appearance of a publication describing the corresponding adverse event was calculated. Cumulative data mining and 1‐ to 2‐year sliding windows produced the most SDRs for recently approved drugs. In the first postmarketing year, data mining produced SDRs an average of 800 days in advance of publications regarding the corresponding drug‐event combination. However, this timing advantage reduced to zero by year 4. The optimum window width for sliding windows should increase with time on the market. Data mining may be most useful for early signal detection during the first 3 years of a drugs postmarketing life. Beyond that, it may be most useful for supporting or weakening hypotheses.

Data Mining in Pharmacovigilance: Computational Cost as a Neglected Performance Parameter

1 Risk Management Strategy, Pfizer Inc., New York, New York, USA 2 Department of Medicine, New York University School of Medicine, New York, New York, USA 3 Department of Community and Preventive Medicine, Department of Pharmacology, New York Medical College, Valhalla, New York, USA 4 School of Information Systems, Computing and Mathematics, Brunel University, London, England 5 Department of Statistics, Rutgers University, Piscataway, New Jersey, USA 6 ProSanos Corporation, Harrisburg, Pennsylvania, USA

A call to incorporate systems theory and human factors into the existing investigation of harm in clinical research involving healthcare products

This is a joint statement from individual pharmacology and pharmaceutical professionals acting in their own capacity, including members of the Alliance for Clinical Research Excellence and Safety (ACRES) and the International Society of Pharmacovigilance (ISoP). By building on the extensive pharmacological and regulatory investigations that already take place, we are calling for a fuller and more robust systems-based approach to the independent investigation of clinical research when serious incidents of harm occur, starting with first-in-human clinical trials. To complement existing activities and regulations, we propose an additional approach blending evidence derived from both pharmacological and organizational science, which addresses human factors and transparency, to enhance organizational learning and continuous improvement. As happens with investigations in other sectors of society, such as the chemical and aviation sector, this systems approach should be seen as an additional way to understand how problems occur and how they might be prevented in the future. We believe that repetition of potentially preventable and adverse outcomes during clinical research, by failing to identify and act upon all systematic vulnerabilities, is a situation that needs urgent change. As we will discuss further on, approaches based on applying systems theory and human factors are much more likely to improve objectivity and transparency, leading to better system design.

Is Earlier Signal Detection Always Better

We read with interest the paper of Lerch et al. [1], which describes a significant yield of true safety signals upon ‘‘resignaling’’, i.e. re-examination of data regarding a drugevent combination that was previously reviewed and considered not to represent a true safety signal. Early detection of true safety signals has been the Holy Grail of pharmacovigilance, and the ability to detect signals early is frequently used as the criterion by which signaldetection algorithms are evaluated and compared. This was the case for signal detection algorithms which made use of spontaneous reports [2, 3], and this criterionwas carried over into subsequent work on algorithms which detect safety signals in electronic health records [4–6]. Time to detection was a key criterion in the recent Observational Medical Outcomes Partnership algorithm competition [7]. The Lerch et al. paper leads us to question this approach, in that a significant number of new safety signals appeared only upon a second assessment with additional data present. Are our signal-detection algorithms now so sensitive that they routinely detect signals before there is adequate accumulated data to properly assess them? We must keep in mind that signal-management is a multi-step process, involving detection, assessment, and decision-making to change a label. If a signal for a given drug-event combination is detected ‘‘too early’’, leading to an inconclusive assessment, it may then go into a monitoring queue where it will eventually re-signal and be reassessed. The overall time from detection until the event appears on the drug label may actually be longer than it would have been if a less-sensitive signal detection algorithm had been used. Reporting patterns may also change over time. Early reports may be sporadic and relatively uninformative. If reporters perceive a pattern of association between a drug and an event and perceive they are contributing to medical knowledge [8], they may report more consistently and provide the information necessary to describe the association to market authorization holders and regulatory agencies. Truly optimum signal detection would identify drugevent combinations at exactly the point where there is sufficient cumulative data to make a proper medical and scientific assessment of whether or not a real adverse drug reaction has been identified—no sooner, no later. Future evaluations of signal-detection algorithms may need to include the total time from detection to label change, accounting for re-signaling and re-assessment. This of course complicates and delays the evaluation process considerably, and surrogate measures may need to be developed which reward early detection, but only up to a point. Certain fruits aren’t ready to be harvested until they have had time to ripen. Does the same principle need to be applied to signal detection in pharmacovigilance?