Researchers trained machine-learning algorithms to read Amazon reviews for hints that a food product would be recalled by the FDA. Christopher Intagliata reports.
The Food and Drug Administration has to recall hundreds of foods every year. Like cookie snack packs with chunks of blue plastic hiding inside, Salmonella-tainted taco seasoning or curry powder laced with lead.
It can take months before a recall is issued. But now researchers have come up with a method that might fast-track that process, leading to early detection and, ultimately, faster recalls.
The system relies on the fact that people increasingly buy foods and spices online. And people tend to write reviews of products they buy online—which are like bread crumbs to food-safety officials sniffing out dangerous products.
The researchers linked FDA food recalls from 2012 to 2014 to Amazon reviews of those same products. They then trained machine-learning algorithms to differentiate between reviews for recalled items and reviews for items that had not been flagged.
And the trained algorithms were able to predict FDA recalls three quarters of the time. They also identified another 20,000 reviews for possibly unsafe foods—most of which had never been recalled. The results are in [JAMIA Open]. [Adyasha Maharana et al., Detecting reports of unsafe foods in consumer product reviews]
The World Health Organization estimates that 600 million people worldwide get sick annually, and more than 400,000 people die from contaminated food. “So having tools that allow us to detect this faster and hopefully investigate and do recalls faster will be useful not just in the U.S. but in other countries around the world as well.” Study author Elaine Nsoesie of Boston University.
She did add one caveat: even recalled products can still get five-star reviews. So stars alone don’t tell the whole sickening story. The proof, unfortunately, may still be in the pudding.
[The above text is a transcript of this podcast.]