Researchers from the Boston University School of Public Health have devised a new artificial intelligence (AI)-driven approach to predicting food product recalls.

According to its developers, the system, which was trained on crowdsourced data from online reviews, predicted food product recalls with 74% accuracy. To do this, the researchers examined almost 1.3 million food product reviews on Amazon.com, and matched over 5,000 of the reviews to products that had been recalled by the FDA between 2012 and 2014. The team then trained a deep-learning AI called Bidirectional Encoder Representation from Transformations (BERT) to locate red flags in customer product reviews.

Relying on the crowdsourced data to train BERT in the identification of unsafe foods, the team categorized approximately 6,000 reviews that contained language associated with FDA recalls including words like “ill,” “sick,” “rotten,” “foul” and “label.” This data was examined along with metadata found in the review titles and star ratings.

Reviews were organized into one of four categories: reviewer became ill or experienced an allergic reaction or there was an error in product labeling; product was expired or tasted/smelled rotten; the review did not contain language suggesting the product was unsafe; or none of the above.

Responsible for nearly 76 million illnesses in the U.S. each year alone, unsafe, contaminated or mislabeled foods can take the FDA considerable time to identify before issuing recalls. However, the BU researchers hope that the AI system will speed up the process of issuing product recalls, thereby reducing the number of associated illnesses.

“Health departments in the US are already using data from Twitter, Yelp, and Google for monitoring foodborne illnesses,” said the study’s senior author, Elaine Nsoesie, assistant professor of global health.

“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,” added Nsoesie.

The research appears in the Journal of the American Medical Informatics Association (JAMIA) Open.