Crowdsourcing Side Effects

Researchers at Microsoft, Stanford and Columbia Universities have developed software that can trawl web searches to find evidence of prescription drug side effects

The team looked at searches from 2010 for an antidepressant, paroxetine; a cholesterol lowering drug, pravastatin; and high blood sugar, a side effect that the two can produce when used in combination.

From a sample of 82 million different searches, they looked for these drugs names as well as for 80 symptom descriptions such as 'blurry vision'.

People who searched for both drugs were twice as likely to also search for such symptoms, compared with those people just searching for one of the drugs. Indeed, 30% did their searches all on the same day.

The US Food and Drug Administration's Adverse Events Reporting System had, in the intervening time, also found this problematic interaction. But that relies on patients raising it with their doctors - and the doctor reporting it. In principle, this sort of 'data mining' could highlight problems long before the conventional system is able to gather enough data, and flag up issues that health officials should take a closer look at.

This sort of data mining is not a new idea. Google Flu Trends has been tracking people's searches for flu-related terms since 2008, and has largely been successful in predicting outbreaks faster than the US Centres for Disease Control. Researchers have also used Twitter data to track the 2010 cholera outbreak in Haiti.

But the system is far from perfect. This winter, Google Flu Trends has massively overestimated peak flu levels in the United States. Experts think that may be because there was greater media coverage of the severe flu season this year, so people were searching for information about it even if they did not have the flu.

The next goal for researchers to refine their models to weed out this sort of signal, and prove that the systems are reliable and timely enough to translate this surveillance into healthcare action - at the moment it has been almost entirely confined to retrospective 'proof of principle' studies. In the near term, it seems that this sort of data mining is going to become a useful support for traditional epidemiological surveillance, but is unlikely to replace it.