You’re feeling unwell, and wonder if the problem might be a side effect of the drugs your doctor just prescribed. So you turn to a search engine and type in your symptoms and the names of the drugs.
It happens every day, and now researchers at Stanford University in California have shown that search logs from millions of web users can be used to detect dangerous interactions between drugs – which are notoriously difficult to predict and monitor.
In 2011, the team mined the US Food and Drug Administration’s database of reported adverse drug events to find that two commonly used drugs – the antidepressant paroxetine and pravastatin, used to lower blood cholesterol – can in combination put patients at risk of developing diabetes.
Many drug side effects never get noticed by doctors, and are not reported to the FDA. So the researchers wondered if web searches by anxious patients might provide a more sensitive tool to find dangerous drug interactions, and teamed up with Microsoft Research to find out.
A spike in searches
Through an add-on to the Internet Explorer browser, millions of web users have consented to let Microsoft collect anonymised data on their searches on Google, Bing and Yahoo. To test whether these logs also contained a signal of the paroxetine-pravastin interaction, the researchers looked at searches by 6 million people in 2010 – before the interaction was discovered – for queries containing both or either drug, and computed how often these occurred in combination with searches for symptoms of high blood sugar, such as “dehydration”, “blurry vision” and “frequent urination”.
Sure enough, there was a clear spike in searches combining the symptoms together with both drug names, over and above the signal for either drug searched alone.
To confirm that the result wasn’t a fluke, the researchers repeated the process for 31 other combinations of drugs known to raise blood sugar, and 31 known not to interact in this way. The rates of false positives and false negatives were similar to those for commonly used medical tests, such as mammography for breast cancer.
Having proved the principle, it should now be possible to go fishing in web search logs for novel side effects, says Lang Li of Indiana University School of Medicine in Indianapolis, who was not involved in the study. “The next stage is to throw in all the drugs and all the adverse drug interactions you can think of and see what comes out,” he says.
The media bias
Still, experience with a tool called Google Flu Trends provides a cautionary tale for putting too much trust in web search data. In previous years, the project has mined search logs to predict the course of the annual US flu epidemic, ahead of surveillance by the Centers for Disease Control and Prevention. But this year, media reports about flu apparently triggered a flurry of searches, causing Google Flu Trends to seriously overestimate the peak of the epidemic.
Nigam Shah, a member of the Stanford team, warns that similar biases could occur with drug interactions. But he’s optimistic that the problem can be solved by combining web search data with the FDA’s adverse events database and another source that Shah has found to be useful to mine for signs of drug interactions – electronic medical records.
“If these three sources individually have a signal, what if we combine them?” Shah asks. “We’re working on systematic ways to do that.”
Journal reference: Journal of the American Medical Informatics Association DOI: 10.1136/amiajnl-2012-001482
If you would like to reuse any content from New Scientist, either in print or online, please contact the syndication department first for permission. New Scientist does not own rights to photos, but there are a variety of licensing options available for use of articles and graphics we own the copyright to.
All comments should respect the New Scientist House Rules. If you think a particular comment breaks these rules then please use the “Report” link in that comment to report it to us.
If you are having a technical problem posting a comment, please contact technical support.