Finding hidden side effects with web search data
Photo: Michel O'Sullivan
Using data drawn from queries entered into Google, Microsoft and Yahoo search engines, scientists at Microsoft, Stanford and Columbia University have for the first time been able to detect evidence of unreported prescription drug side effects before they were found by the US Food and Drug Administration's (FDA) warning system.
Using automated software tools to examine queries by 6 million internet users taken from web search logs in 2010, the researchers looked for searches relating to an antidepressant, paroxetine, and a cholesterol-lowering drug, pravastatin. They found evidence that the combination of the two drugs caused high blood sugar.
The study, which was reported in the Journal of the American Medical Informatics Association on Wednesday, is based on data-mining techniques similar to those employed by services like Google Flu Trends, which has been used to give early warning of the prevalence of the sickness.
Currently, the FDA asks physicians to report side effects through a system known as the Adverse Event Reporting System, but its scope is limited by the fact that data is generated only when a physician notices something and takes the initiative to report it.
The new approach is a refinement of work done by the laboratory of Russ B. Altman, the chairman of the Stanford bioengineering department. The group had previously explored whether it was possible to automate the process of discovering "drug-drug" interactions by using software to hunt through the data found in FDA reports.
The group reported in May 2011 that it was able to detect the interaction between paroxetine and pravastatin in this way. Its research determined that the patient's risk of developing hyperglycemia was increased compared with taking either drug individually.
The new study was undertaken after Altman wondered whether there was a more immediate and more accurate way to gain access to data similar to what the FDA had access could get.
With the aid of Microsoft researchers, he acquired anonymised data taken from a software toolbar installed in web browsers by users who had allowed their search histories to be collected. The Microsoft researchers were able to look at 82 million searches for drug, symptom and condition information.
The researchers first identified individual searches for the terms paroxetine and pravastatin, as well as searches for both terms, in 2010. They then computed the likelihood that users in each group would also search for hyperglycemia as well as roughly 80 of its symptoms – words or phrases such as "high blood sugar", "blurry vision", "frequent urination" or "dehydration".
They were able to determine that people who searched for both drugs during the 12-month period were significantly more likely to search for terms related to hyperglycemia, compared with those who searched for just one of the drugs. (Approximately 10 per cent, compared with 5 per cent and 4 per cent for just one drug.)
They were also able to determine that people who did the searches for symptoms relating to both drugs were likely to do the searches in a short time period: Thirty percent did the search on the same day, 40 per cent during the same week and 50 per cent during the same month.
"You can imagine how this kind of combination would be very, very hard to study given all the different drug pairs or combinations that are out there," said Eric Horvitz, a managing co-director of Microsoft Research's laboratory in Redmond, Washington.
The researchers said they were surprised by the strength of the "signal" that they were able to detect in the millions of web searches and argued that it would be a valuable tool for the FDA to add to its current system for tracking adverse effects.
"There is a potential public health benefit in listening to such signals," they wrote in the paper, "and integrated them with other sources of information."
The researchers said they were now thinking about how to add new sources of information, like behavioural data and information from social media sources. The challenge, they noted, was to integrate new sources of data while protecting individual privacy.
Currently the FDA has financed the Sentinel Initiative, an effort begun in 2008 to assess the risks of drugs already on the market. Eventually, that project plans to monitor drug use by as many as 100 million people in the US. The system will be based on information collected by health care providers on a massive scale.
The New York Times