FDA-funded study shows limits of mining Twitter for adverse event reports

The steady flow of posts about adverse events on social media provide an interesting opportunity for the FDA to improve postmarketing safety surveillance. Yet an FDA-funded study has found that mining the data for insights is difficult, with humans still better equipped than machines to decipher chatter on Twitter ($TWTR).

A team of researchers from Harvard Medical School and other academic centers examined more than 60,000 tweets, of which 4,401 were manually categorized as resembling adverse events. Over the same period as the tweets were gathered, the FDA recorded 1,400 incidents in its Adverse Event Reporting System (FAERS). The number of tweets compared to official reports highlights the potential of mining social media for adverse events, but also hints at why doing so will be technically demanding.

For widely used drugs the number of tweets is likely to overwhelm FDA capacity to manually review the posts. Computers could potentially help, but the study authors doubt whether today's technology is up to the task. To illustrate the challenge, the paper singles out the following tweet: "Xeljanz was the best but ate a hole in my stomach." The statement sounds like an exaggeration, but the FDA label for Pfizer's ($PFE) Xeljanz lists tears in the stomach as a possible side effect.

If the post was discussing aspirin, it could be dismissed as an exaggerated report of routine stomach discomfort. But in the case of Xeljanz the comment could be interpreted more literally and therefore more seriously. While the ability of computers to interpret and contextualize text is improving, the paper concludes that classification by machines followed by curation by humans is the best approach for now. This workflow would still require significant human resources.

The authors--who include FDA's recently hired chief health informatics officer, Taha Kass-Hout--also advise against wholesale importing of social media posts into postmarketing safety databases. Instead, they recommend using tweets in conjunction with other sources to gather ideas about what possible adverse events are worth taking a closer look at in epidemiologic studies.

- read the paper