Further improvements to the sensitivity and specificity of the method include stratifying by age and sex; examining serious or fatal reactions only. (The proportion of reactions that are fatal, stratified by age is itself a potential signalling method.) Where a drug has a well-known ADR that is reported very frequently, such as gastrointestinal (GI) bleeding with non-steroidal antiinflammatory drugs (NSAIDs), this will distort the PRR for other reactions with that drug. The best approach is then to remove the known reactions from the totals for that drug and the database as a whole and recalculate the PRR for all other reactions with that drug. This is simple to do on an ad hoc basis but is more difficult to implement in an automated way.
The comparison used need not be the entire database. It is possible to use PRRs within drug classes or indications so that the comparator is all drugs in that class or those used for a particular indication.
The expected number of reactions could also incorporate prior beliefs about the ADR profile, using a fully Bayesian method. (The approaches used at the FDA and WHO do not incorporate prior beliefs.)
The grouping of terms used in the medical dictionary for the database is an important feature. Little empirical study of the effect of choosing different levels in the hierarchy of terms has been done. In most instances, the grouping is at "Preferred Term'' (PT), which is a relatively low level. There are a large number of medical terms at this level, so that the numbers for any particular combination of drug and reaction can be small. This can lead both to the general statistical problem of multiplicity, with many possibilities for signals, and to instability in the PRR based on small expected numbers.
It is possible to use a two-stage process—using, say, SOC to screen for raised PRRs, then to reexamine the PRRs using PTs within the SOC where the PRR was raised. The automation of this process is possible in principle, but has not been done yet. An alternative is to use an intermediate level within the hierarchy—a "High Level Term'' (HLT), for example. This has the advantage of being a single stage process and avoids the use of too many terms, reducing the problems of multiplicity and small expected numbers.
The use of the method in general is easiest within a large database that contains a wide range of drugs, but it can be used within a pharmaceutical company database. Here, the potential for incorporating prior beliefs is at its greatest. A further possibility for companies is to use the proportions of reactions from the FDA database, which is publicly available, to calculate expected numbers for their own drugs. Other regulatory databases are not yet publicly available but increasing transparency may change this in the future.
Was this article helpful?