FDA Moves Forward With Machine Learning AI Pilot for Import Screening

FDA is moving forward with a pilot program to test the use of artificial intelligence in its import screening, FDA Commissioner Stephen Hahn said Aug. 31 in a blog post on the agency’s website. FDA has successfully completed its first phase proof of concept that it began in spring 2019 (see 1903220030) to test AI machine learning technologies in import screening for seafood, and is now working to begin a second phase to test the concept in the field.

As part of the second phase pilot, the machine learning model “will be applied to the screening methods used to help FDA staff decide which shipments to examine and will then provide information about which food in the shipment to sample for laboratory testing,” Hahn said. “We will then compare the results to the recommendations made by our current system.”

In the proof of concept, FDA had applied the machine learning screening tool to “years of retrospective data from past seafood shipments that were refused entry or subjected to additional scrutiny, such as a field exam, label exam or laboratory analysis of a sample,” Hahn said. “This gave us an idea of how much our surveillance efforts might be improved using these technologies.” The results were “exciting, suggesting that this approach has real potential to be a tool that expedites the clearance of lower risk seafood shipments, and identifies those that are higher risk,” he said.

“The proof of concept demonstrated that [machine learning] could almost triple the likelihood that we will identify a shipment containing products of public health concern,” Hahn said. FDA sees the pilot as a step toward deploying AI across all of its regulated products, giving the agency the ability to “untether the knowledge we need from the huge volume of data we have from screening millions of import shipments every year,” Hahn said.

“The FDA has a massive amount of data about these shipments and about the companies that are producing and processing the food, offering it for import, and selling it in the U.S. marketplace,” Hahn said. “In fact, every year the FDA collects tens of millions of data points on imports alone, and we screen all the data associated with every shipment of food against the information in our internal databases. One of the major goals of our pilot is to assess the ability of [AI machine learning] to more quickly, efficiently, and comprehensively take advantage of all the data and information residing in our systems,” he said.

“We believe that we can use the knowledge that [machine learning] provides to know where best to concentrate our resources to find potentially unsafe products. In addition to improved import surveillance resources, the intelligence that [machine learning] can extract from the stores of data the FDA collects can also inform decisions about which facilities we inspect, what foods are most likely to make people sick and other risk prioritization questions,” Hahn said.