Medical data mining can predict adverse drug events

Researchers for the RAND Corp. have determined that it's possible to mine medical literature to find out whether studies involving particular medications can predict patient harm from those drugs. Their findings recently were published online in the Journal of the American Medical Informatics Association, reports InformationWeek Healthcare.

The researchers used an algorithm known as a "statistical document classifier" to search PubMED for MEDLINE citations published between 1949 and September 2009 that mentioned one of 38 drugs and one of 55 adverse drug events. They found 9,100 articles that met their criteria.

Among other things, the researchers turned up evidence of heart disease caused by Vioxx in 2001 studies. In articles published in 2002, they found evidence of similar side effects of Celebrex, a related drug. Merck recalled Vioxx in 2004, and Pfizer recalled Celebrex a few months later.

Overall, the RAND model enabled the researchers to detect 54 percent of "detected FDA warnings" about particular drugs in the literature that existed before those alerts were issued.

"Results from large-scale literature retrieval and analysis [literature mining] compared favorably with and could complement current drug safety methods," the RAND team concluded.

The FDA already uses an Adverse Events Reporting System to detect drug harms after prescription medications have been approved. But analyses of the data from this voluntary reporting system tend to be inaccurate, according to InformationWeek Healthcare.

To learn more:
- read the InformationWeek Healthcare story
- here's the study abstract in JAMIA