Data mining is intended to alert the observer to unusual relationships within a data set. It is essential to understand that in pharmacovigilance, what is reported and contained within the data set does not represent the true epidemiology of adverse reactions to medicines. There is the very well-known problem of under-reporting, but more than that, many countries ask health professionals to be selective in their reporting to cut down the "noise". One problem with data mining is the temptation to turn it into data dredging. There is a difference: data mining uses objectively predetermined (if flexible) logic to examine relationships in data transparently. Data dredging is based upon a series of prejudiced queries which might imbue chance relationships with plausibility, and in which a strict logic or strategy is not followed.
In the past, it has seemed reasonable for pharmacovigilance experts to reduce their work load and avoid having to see multitudes of reports of more trivial or well-known adverse reactions, but this has both health and methodological consequences. It is often forgotten that "serious adverse and unexpected'' reactions can be preceded by less serious phenomena. The best known is the xerophthalmia related to practolol being the harbinger of sclerosing peritonitis. Also, the persistent reporting of a well-known (to experts) adverse reaction-product combination can be important since it may indicate that practitioners in the field are concerned about it for some practical reason. The reasons may be that they see the reaction more frequently than they think they should, that there is something unusual about the duration or severity, or that there are systematic errors associated with the use of the product which lead to problems (similar confusing labelling of different products, for example).
Data mining should allow for much easier and useful handling of large amounts of information. Since the "triaging" of information is done automatically, there is no longer any need to specify that only serious and unexpected reactions need be reported. Indeed, data mining in pharma-covigilance will function better for us if there is a large amount of "ordinary" adverse reaction in formation to serve as the background. If we just record the serious and unexpected, only the more serious and unexpected will stand out, progressively. This slow shift of emphasis would be deleterious for public health.
Data mining has its main future in the detection of complex patterns in the data. It is possible that, if doctors reported all the medicinal product safety issues that concern them, we would be able to identify some issues of use and poor use of medicines that could be addressed (Edwards and Aronson, 2000).
Data mining, then, is proving to be a useful tool. Its full potential has not yet been reached, and it may be that some of the current drug regulations and attitudes may need to be reconsidered as its use becomes more widespread. In spite of its potential as the primary search tool in pharmacovigilance, it is clear that its use must be accompanied by the wise interpretation of the information. Since no database is representative of what truly happens, other observations, monitoring and epidemiology must continue to be used in a complementary way. Only by the interactive interpretation of findings using different observational methodology are we likely to even approach the truth.
Was this article helpful?