By applying analytics to medical journal articles, Rand researchers were able to uncover dangerous side effects before regulators' existing systems could.
(click image for larger view)
Slideshow: 12 Innovative Mobile Healthcare Apps
Data mining of the medical literature could help uncover drug side effects before they cause serious harm to patients, a new study suggests.
Researchers from Santa Monica, Calif., think tank Rand surmised that a review of published studies could help regulators, like the U.S. Food and Drug Administration, spot dangerous uses of drugs earlier and prevent situations like the 2004 recall of rofecoxib--sold under the brand name Vioxx--following revelations that the arthritis drug could increase the risk of heart attack and stroke.
"Regulatory agencies and drug safety researchers may be able to use these techniques to improve decision-making about drug safety," Rand CTO Siddhartha Dalal and researcher Dr. Kanaka D Shetty wrote in a new article published in the Journal of the American Medical Informatics Association.
Shetty and Dalal developed an algorithm to mine PubMed, the free portal for Medline, the National Library of Medicine's bibliographies of medical literature. They looked for mentions of least one of 38 specific drugs and 55 adverse effects, then determined the relevance of the articles and forecast expected rates of adverse effects (AEs). While others have processed large databases to find evidence of side effects, including the FDA's Adverse Events Reporting System, such analyses have tended to be inaccurate.
"Prior work suggested that up to 98% of search results are irrelevant to [adverse effects]; including such articles would skew results toward false positive drug-AE linkages," they wrote.
With this in mind, the Rand team initially found more than 119,000 articles published between 1949 and September 2009, then excluded articles that didn't test whether a drug caused an AE. They ended up analyzing about 9,100 articles, plus a wider set of articles that took into account side effects from classes of drugs rather than individual compounds.
"We aimed to prototype a process for collecting and analyzing relevant literature while minimizing false positive and false negative drug-AE associations," Shetty and Dalal said. "We first improved the data collection process by excluding irrelevant articles using supervised statistical learning techniques that automatically filter citations using Medline indexing terms. We then used statistically valid machine learning algorithms ... to identify significant drug-AE links from the several thousands of such pairs while controlling the overall probability of errors.
"Finally, we evaluated whether the literature-mining process might have detected drug-AE associations that were the subjects of FDA warnings. We used the entire literature in some analyses and simulated prospective analyses by restricting the data to literature available prior to the warnings."
They ended up finding evidence of heart disease caused by Vioxx as early as 2001, from literature published by 2002, and Merck pulled the drug from the market in 2004. "We dramatically improved [positive prediction] and sensitivity rates with respect to major known drug-AE associations by applying successive filters and incorporating drug class effects," the researchers wrote.
"Regulatory agencies and drug safety researchers may be able to use these techniques to improve decision-making about drug safety," Shetty and Dalal concluded. "Furthermore, we detected numerous associations prior to FDA warning, suggesting that literature mining did not simply provide a lagging indicator of widely known drug-AE associations."
Join us for a webcast and find out the real-world issues facing healthcare providers as they implement BI and analytics. It happens May 26. Find out more. (Free with registration.)