Researchers from Boston University School of Public Health wants to use artificial intelligence to find and identify dangerous food products being sold on Amazon.
In a study, the team trained deep learning AI to peruse customer-submitted reviews on the e-commerce site and predict which items will be recalled by the Food and Drug Administration or FDA.
Flagging Potentially-Harmful Food On Amazon
To do this, the researchers collected a total of 1,297,156 consumer-submitted reviews of food products sold on Amazon.com. They linked these reviews to product recalls made by the FDA from 2012 to 2014.
The researchers asked volunteers to comb through 6,000 reviews that contained words and terminologies that the FDA used to justify product recalls in the past. This includes words such as "sick," "rotten," and even label."
The volunteers classified these reviews into four different categories: the reviewer got sick, had an allergic reaction, found an error in the product's labeling; the product looked or tasted bad, was expired, needed inspection; the reviewer made no claims that the product is potentially unsafe; none of the above.
Using the above data, the researchers trained a type of deep learning AI called Bidirectional Encoder Representation from Transformation, or BERT, to identify which foods have been recalled by the FDA.
The researchers reported that BERT was able to correctly flag recalled products with 74 percent accuracy. Moreover, the AI found an additional 20,000 products that have not been officially recalled by the FDA but match the criteria.
Helping The Authorities Identify Unsafe Food Products
The researchers explained that identifying and then investigating a potentially dangerous food product can take months before federal authorities issue a recall. Often, manufacturers voluntarily recall their products after many people have gotten sick.
Using AI to sort through consumer reviews on popular e-commerce sites can help the FDA to find unsafe food products and issue a timely recall.
"Health departments in the US are already using data from Twitter, Yelp, and Google for monitoring foodborne illnesses," stated Elaine Nsoesie, an assistant professor at Boston University of Public Health and one of the authors of the study. "Tools like ours can be effectively used by health departments or food product companies to identify consumer reviews of potentially unsafe products, and then use this information to decide whether further investigation is warranted."
The study was published in the Journal of the American Medical Informatics Association on Monday, Aug. 5.
