Trimethoprim-induced hyperkalaemia – lessons in data mining

The principle concern of phamacovigilance is the discovery of adverse events (AEs) that are novel in terms of their clinical nature, severity and/or frequency as early as possible after marketing with minimum patient exposure. Spontaneous reporting systems (SRS) play a central role in this enterprise and constitute the main data source for gleaning ‘signals’ of previously unrecognized AEs. With ever increasing volumes of postmarketing safety data, research is focusing on data mining algorithms (DMAs) that can search extremely large SRS databases for disproportional statistical dependencies between drugs and AEs [1–4]. Given sufficient correlation between the observed statistical dependencies and demonstrable causal drug–event associations (DEAs), DMAs could enhance early detection of signals of novel AEs.

The most commonly studied DMAs include forms of simple disproportionality analysis [e.g. proportional reporting ratios [2] (PRRs)] or methods incorporating statistical adjustments based on Bayesian inference, and extensive covariate stratification and additional data manipulation [e.g. multi-item gamma-Poisson shrinker [3] (MGPS)]. A potential advantage of certain Bayesian methods is a reduced volume of signals from ‘noise’ reduction [4]. However, this may not be entirely beneficial if important causal DEAs are filtered out by Bayesian methods.

A crucial question is the comparative performance of the various methods and their proper positioning relative to standard signal detection techniques. The answer requires an adequate understanding of their performance characteristics and factors influencing performance. Studying these methods in a variety of diverse settings will add to the cumulative knowledge of their strengths and weaknesses and promote optimal choice and deployment of signal detection strategies.

Trimethoprim-associated hyperkalaemia nicely illustrates certain relevant considerations in data mining such as the potential performance differentials between simple disproportionality analysis and Bayesian methods with respect to early signalling capability using commonly cited thresholds. This AE typifies one type of AE that would hopefully be susceptible to early detection using data mining [5]: (i) it is a medically serious, and possibly life-threatening AE; (ii) its recognition by clinical observation was delayed, probably due to the lack of an obvious mechanistic link between drug and event (prior to recognition of trimethoprim (TMP)'s amiloride-like structure and activity at the distal nephron); (iii) the association is significantly complicated by co-morbid illnesses, nutritional status, and comedications; (iv) it is not always present on lists of ‘designated medical events’ that by virtue of their seriousness, rarity and high drug-attributable risk, have very low signal thresholds in traditional pharmacovigilance practice.

Using data from the Food and Drug Administration (FDA) Adverse Event Reporting System (AERS) through the first quarter of 2003, both methods were applied using previously published threshold criteria and configurations [2, 3] to the suspect DEAs of trimethoprim and trimethoprim-sulfamethoxazole-associated hyperkalaemia. A signal was generated with PRRs by three reports (4 years before the first published case report and almost 10 years prior to wider recognition). Only with the accumulation of 56 reports was a signal generated with MGPS a full 15 years after the signal generated by PRRs and long after the first published case report.

This is but one example of how PRRs have the ability to highlight important DEAs that may escape detection by Bayesian methods as well as highlight adverse events earlier when commonly cited thresholds are used. The cost of this enhanced sensitivity is an increased volume of false-positive signals not reflective of causality that may require additional triage criteria for practical implementation. Given the predictable trade-off between sensitivity and specificity, future research should compare various threshold criteria for each method when used as a component of a comprehensive pharmacovigilance programme with a wide variety of drugs and events in order to determine their optimum deployment.