The AI That Can Save Millions From Unsafe Food

If there is one thing that AI has proven time and time again that it excels at, it’s using massive data records to deduce patterns and make surprisingly precise predictions. Recently, a research group from Boston University School of Public Health has managed to create a set of machine learning algorithms that can spot unsafe and hazardous food items by studying public food reviews on Amazon. It then tags these items for review and potential recall from the marketplace altogether.

A few days ago in the American Medical Informatics Association Journal, a study was published and detailed every procedure that the scientists carried out in order to develop the AI. Almost 1.3 million reviews of food items were gathered by the group using a separate programme, and then around five thousand of them were connected to items that had already been recalled by the FDA (USA Food and Drug Administration) in 2012, 2013 and 2014. This meant that the AI could differentiate from safe and found unsafe products and could therefore detect patterns in the nature of the reviews.

The FDA is a federal agency of the US Department of Health and Human Services. Source: Federal News Network

The FDA is a federal agency of the US Department of Health and Human Services. Source: Federal News Network

The researchers decided to make use of a machine learning AI called “Bidirectional Encoder Representation from Transformations”, also known as “BERT”. However, before they could train the AI on the reviews, they had to categorise reviews on unsafe products that the AI might get confused with. Real humans ended up moving around six thousand reviews into four separate classes based on what the reviewer was saying about the food item. Unfortunately, when setting up any AI human workers must do a lot of the work themselves, no matter how tedious.

However, this effort certainly paid off. In the end, making use of all the reviews sorted or otherwise, the BERT AI could faithfully classify the food products that had been recalled by the FDA with a solid 74% accuracy. The most interesting part was that the artificial intelligence programme also was able to spot about 20,000 unrelated food items that have never been recalled by the FDA.

There are undoubtedly some problems with using public reviews from a website such as Amazon, as reviews can be untruthful or misleading and most are not from professional food critics. Nonetheless, such an easy and effective instrument will be a useful tool in any regulator’s toolbox, especially for slow moving bodies such as the FDA. Even if only used as evidence of danger, this could be utilised to pressure food manufacturers into changing their practices for the better without government departments having to use their limited resources.

The journal study can be found in full here.

Thumbnail source: geneticliteracyproject.org

ResearchEdward Bristow