AI could help food systems detect pathogens, fraud, and contamination faster

by · News-Medical

From detecting Salmonella to flagging risky food suppliers, a new review shows how AI is moving food safety research toward faster, more predictive monitoring

Review: Artificial intelligence in food safety. Image Credit: TSViPhoto / Shutterstock

The review's findings document a rapid rise in AI research and its reported use in published food safety studies, alongside a shift toward advanced deep learning models. Ultimately, the review underscores AI’s potential to proactively support safer, more resilient global agrifood systems and strengthen food safety decision-making.

Background

Ensuring food safety is a major global challenge with repercussions for public health, economic stability, and food security. Traditionally, safeguarding food intended for human consumption has largely relied on reactive measures: testing food samples after production or investigating outbreaks after consumers fall ill.

However, modern agrifood networks generate volumes of data that traditional manual inspection methods are incapable of processing efficiently. To address this conventional limitation and ensure a safe and validated food supply, experts are increasingly turning to artificial intelligence (AI).

While classical statistics has long monitored risks, AI introduces tools such as machine learning, which extracts features from data to predict outcomes, and deep learning, which automatically interprets raw datasets.

The recent surge in interest in leveraging AI in food science is best exemplified by the number of peer-reviewed studies published each year. In 2012, only one study focused on the application of AI in food safety; by 2023, that number had increased to 46. Yet, until now, a cohesive global map outlining exactly how these algorithms are being studied and applied across food safety research has remained missing.

About the review

The present review aimed to synthesize this expanding scientific landscape and provide a roadmap for future research in the field. The review initially screened 783 candidate publications from the SCOPUS database.

To streamline the screening process, the review deployed an active learning software tool called “ASReview”. ASReview is an ML tool that sequentially ranks papers based on predicted relevance. Notably, the ML tool was able to refine its choices with each researcher's input. Using ASReview, the researchers screened 434 records at the title and abstract level before conducting a full-text evaluation.

Following full-text evaluation, 161 primary research articles and peer-reviewed conference proceedings published up to April 2024 were selected for final analysis. These publications were classified based on their research domain, implementation context, data collection methods, other methodologies, and the specific AI architecture employed.

The review categorized studies across research domains, including microbiological hazards, chemical contaminants, food authenticity, foodborne disease outbreak surveillance, and broader food safety issues.

Review findings

The review’s analyses revealed that AI is heavily concentrated in certain sectors, with microbiological hazards leading at 35% of the reviewed literature. Within microbiology, 59% of studies used AI to augment conventional laboratory testing. For example, one study paired an "electronic nose" sensor with classification algorithms to identify Salmonella, achieving 85% to 100% accuracy. In another, a random forest model predicted disease endpoints from untagged Salmonella genetic sequences with 87% accuracy.

After microbiological hazards, chemical contaminants were the next most common research domain, accounting for 25% of the reviewed literature. Herein, AI was predominantly used to detect heavy metals or pesticides non-destructively. Food fraud and authenticity accounted for 17% of the papers, highlighting applications like scanning electronic invoices to flag suspicious oil manufacturers.

Separately, AI’s application in disease surveillance was also shown to offer potential advantages over conventional methodologies. One system used anonymized smartphone search and location data to identify contaminated venues, proving more than 3 times as effective as traditional investigations in that study.

Finally, the review revealed a significant increase in the use of deep learning algorithms in recent research, rising from 22% of papers in 2019 to 43% by 2023.

Conclusions

The present review highlights that while AI promises to improve food monitoring, significant hurdles must be overcome before its potential can be fully realized. Chief among these is severe class imbalance, as the vast majority of food safety data reflects safe, low-contamination environments, making it difficult for algorithms to recognize rare, high-risk anomalies. Furthermore, data privacy and proprietary restrictions often prevent open data sharing.

The authors also noted that model performance could not be systematically compared across studies because many datasets contained few positive cases and sparse predictor combinations. In addition, the review was limited to Scopus-indexed literature and may underrepresent commercial or manufacturing applications where findings are not published in peer-reviewed journals.

Moving forward, the authors emphasize that emerging solutions like explainable AI, which demystifies how models make decisions, and decentralized federated learning will be vital. Embracing these innovations could help shift food safety from a largely reactive system to a more predictive, transparent, and data-informed monitoring approach.

Download your PDF copy by clicking here.

Journal reference: