AERS spider: an online interactive tool to mine statistical associations in Adverse Event Reporting System.
Grigoriev I, zu Castell W, Tsvetkov P, Antonov AV. Pharmacoepidemiol Drug Saf. 2014 Aug;23(8):795-801.
Principles to detect signal
Proportional reporting ratio (PRR)
Each attribute, a drug, a reaction, an indication or gender, thus, defines a subpopulation of objects within overall population of objects. The commonly used proportional reporting ratio (PRR) to detect signal in AERS database (Evans et al., 2001), compare the frequency of an Adverse Reaction (AR) in subpopulation of reports defined by a drug and the frequency of the AR in some general population of reports (see figure 1). The general population (further referred to as background) is either the entire population or specific one, like, "DIABET". The role of background population is similar to the role of control group in clinical trials. If the frequencies of AR in the drug subpopulation and background subpopulation are substantially different then a common conclusion is that administration of a drug increase the risk of AR.
Figure 1. Proportional reporting ratio (PRR). The frequency of the Adverse Reaction (AR) in drug subpopulation is defined as a proportion of reports with AR. The frequency of the Adverse Reaction in background is defined as a proportion of reports with AR. Please note that to compute background frequency of AR the drug subpopulation of reports is excluded (shown by blank space). Proportional reporting ratio (PRR) compares the frequencies of AR in drug and background populations.
There are multiple subgroups of reports, defined by various attributes (indication, age group, other drugs), with incomparably high risk of various Adverse Reactions (figure 2). These risk factors, referred further as "mask" factors, may be disproportionately distributed between drug and background population. This makes the prior risk of AR (the risk of adverse outcome before drug usage) in both populations incomparable. The blind use of PRR values without accounting for this fact would inevitably lead to ridiculous conclusions.
Figure 2. "Mask" factors (shown as Factor1, Factor2, Factor3 and Factor4) are subgroups (indications, other drugs, age groups) with incomparable high prior risk of AR. The factors affect substantially the estimate of frequency of the AR either in drug subpopulation (Factor1, Factor3) or in the background population (Factor2, Factor4).
AERS spider computes all potential "mask" factors. By definition, factors (indication, age group, other drugs) with PRR > 3 of association to the Adverse Reaction are considered as potential "mask" factors. These factors are reported in the table 2 and the user can mark them to remove from consideration in the next step (figure 3). PRR for the association of interest is recomputed (accounting for removed factors) as well as the list of the new "mask" factors is updated. The procedure is supposed to stop when no suspicious "mask" factors remains unresolved.
Figure 3. Cleaning Signal: "Mask" factors are removed from consideration (based the user expert opinion). Both frequencies of AR in the drug and the background populations are recomputed accounting for removed factors (shown as blank space).
Risk cofactors may not be directly associated to the Adverse Reaction. However, these factors are asymmetrically distributed between patients who administered the drug with AR and the patients who administered the drug without AR (see figure 4). In other words, these factors either over or under represented in reports with reaction in a drug subpopulation and may increase/decrease the risk of the AR if being co administrated with the drug of interest.
Figure 4. Risk cofactors are factors unequally distributed between patients who administered the drug with AR and the patients who administered the drug without AR. Cofactor 1 is equally distributed. Cofactor 2 is underrepresented in "the drug with AR" subpopulation. Cofactor 3 is overrepresented in "the drug with AR" subpopulation.
In many cases it is more natural to use not the entire report population as background but reports from a more specific large subpopulation where the prior risk of the Adverse Outcome is more adequately modeled. For example, to explore the risk of fracture in relation to "calcium" administration one need either to select as background "OSTEOPOROSIS" reports or to use all reports excluding the indications with "Bone Density" problems.