Researchers developing new automated methods for earlier detection of drug adverse side effects

Published October 08, 2013 by Sarah Kelly

Trevor Cohen, MBChB, PhDPrescription drugs are meant to help patients suffering from illness and disease, but some drugs, like Vioxx, produce unintended side effects after being released into the marketplace.  

“These side effects are not always discovered in clinical trials because individuals with idiosyncratic responses may not be represented in the trial population,” said Trevor Cohen, MBChB, PhD, associate professor at UTHealth School of Biomedical Informatics. “After the drugs are released into the marketplace, new side effects may emerge because more people are taking the drug, and the implications of a side effect that was observed previously may become clearer.”

According to the New York Times, about 25 million Americans took Vioxx in the course of five years, and in 2004, the drug was pulled off the market because evidence showed that it doubled the risk of heart attack, stroke and death.

So, what can be done to mitigate the fatalities and illnesses caused by unintended drug side effects?

“We’re proposing a new automated methodology that will identify plausible drug event pairs based on patient data in electronic health records and information from the biomedical literature,” said Cohen, principal investigator for the project. “Currently, the FDA collects side effect reports and has provided spontaneous reporting systems that physicians can use to report potential side effects, but it has been shown that reports for these systems are biased. People tend to report things that they already know are side effects, and other problems aren’t necessarily reported.

“The novel approach that we’re developing has the potential to detect side effects much earlier than existing methods.”

The team of researchers, funded by the National Library of Medicine through 2016, will utilize patient data from electronic health records that are housed in the UTHealth data warehouse and maintained by Elmer Bernstam, MD, MSE, SBMI professor and associate dean for research. The patient data will be paired with information extracted from the biomedical literature in MEDLINE.

“Our aim is to automatically identify those associations occurring in the data warehouse that are biologically plausible,” said Cohen. “The system will “read” clinical notes in the data warehouse to find side effects and drugs. Then, it will report the number of times each of these drugs and side effects occurred in the electronic health records to a statistical model, which is also automated.

“Afterward, we project these drugs and potential side effects, as well as tens of millions of assertions extracted from the biomedical literature, into a hyper-dimensional geometric space. It is here that analogical reasoning occurs to identify plausible connections between the identified drugs and side effects.”

The biomedical literature and patient data will be processed using natural language processing systems that were developed by Thomas Rindflesch, PhD, creator of a National Library of Medicine system called SemRep, and Hua Xu, PhD, SBMI associate professor, director of the Center for Computational Biomedicine and the Robert H. Graham Professor in Entrepreneurial Biomedical Informatics and Bioengineering. Other collaborators include SBMI PhD student Ning “Sunny” Shang, MS; Peter Davies, MD, PhD, professor, Alkek Chair and director of the Center for Translational Cancer Research at Texas A&M Health Science Center; and Peng Wei, PhD, an assistant professor and biostatistician at the UT School of Public Health.

Current methods report drug side effect associations to an expert who determines if the side effect is plausible based on the current understanding of the drug and its side effects. To reduce the human effort required, the system being created by Cohen and his team will automatically determine the plausibility of an association between a drug and a side effect based on knowledge extracted from the biomedical literature and information derived from electronic health records. However, their system won’t fully eliminate the involvement of people in the decision-making process, especially as it pertains to regulations.

“What the system will provide at the end of the day is a list of what it determines are the most plausible side effects based on the statistical associations and evidence from the literature that supports these assertions,” said Cohen. “The implication is that we’ll be able to detect potential side effects earlier. Imagine the number of lives that would have been saved if the side effects of these drugs that are now off the market were found a year earlier.”