Data Mining Part Deux: A New Approach to Drug Safety Analysis via the Internet
Data mining from spontaneous adverse event reports submitted to Medwatch and included in FDA’s Adverse Events Reporting System (AERS) database has become an important activity for both FDA and the drug and device industries. Such approaches, used routinely in the world of FDA-regulated products, allow connections between specific drugs and specific adverse events to be identified based on disproportionate associations. Somewhat more recently it has been acknowledged that drug-related problems can be recognized from analyses of electronic medical records (EMR), providing a view into much of the “iceberg” that is known to be poorly sampled from spontaneous event reporting processes. The approach via EMR can be quite time-consuming and costly, however, typically involving considerable effort related to protocol and software development, approval by Institutional Review Boards, etc. A recent publication titled “Web-scale pharmacovigilance: listening to signals from the crowd” from the Journal of American Medical Informatics Association has highlighted yet another approach.
Internet search approaches have previously demonstrated the ability for seasonal influenza health tracking. In this latest example, White and colleagues were able to use evaluation of common search engine activities with Google, Yahoo and Bing to identify the association of a specific drug interaction (paroxetine-pravastatin) and hyperglycemia. The premise of such a research exercise was that patients are likely to learn about the drugs they are taking to help explain symptoms they are experiencing. Using a medically-appropriate list of 78 hyperglycemia-related search terms, internet queries for both drugs and also for the symptoms of hyperglycemia can be quantified from all internet searches. Such searches involving both drugs are distinguished from searches involving either drug alone to highlight potential drug interactions. The figure reproduced to the left from the publication identifies the search subpopulation of interest.
Analysis of the search results demonstrated that people who searched for both paroxetine and pravastatin during calendar year 2010 were approximately twice as likely to have also performed searches on the terms associated with hyperglycemia than those who searched on hyperglycemia and only one of the drugs. Confirmation of this approach was presented by a web-search analysis for a number of other drug pairs already known to interact and cause hyperglycemia (true positive associations) and a number of randomly chosen drug pairs not associated with hyperglycemia (true negative associations): internet searches much more commonly involved the true positive drug interactions.
Why is this of interest? First, these types of data don’t involve the same kind of reporter bias that is known to exist for routine post-market adverse event reporting via Medwatch. Second, searching by web users is much more common than adverse events reporting to FDA; that is, the source data are much more representative of problems in clinical medicine. And third, evaluation of logged Web search activities does not involve the unavoidable latency associated with drug safety surveillance from spontaneously reported adverse events included in a centralized database. Case in point, according to the authors the paroxetine-pravastatin drug interaction associated with hyperglycemia was not known at the time the 2010 web log data were collected.
Posted by Bob Roth, Vice President and Worldwide Medical Director. For more information, please contact Bob at email@example.com.