Scientists Analyzed 28,724 Food Safety Alerts and Found a Surprising Limit to AI

By Hugo Francisco de SouzaReviewed by Susha Cheriyedath, M.Sc.Oct 21 2026

Can thousands of past food safety alerts reveal which cross-border warning links will emerge next, or are sophisticated AI models less predictive than they appear?

Study: Evaluating directed and multiplex network representations for prediction of emerging contamination pathways in food safety surveillance. Image Credit: Dusan Petkovic / Shutterstock

In a recent study published in the journal Scientific Reports, researchers evaluated whether network-based machine learning (ML) models can predict emerging international food safety notification links. The investigation analyzed 28,724 notifications from the European Union's Rapid Alert System for Food and Feed (RASFF; 2019 to 2025) to assess whether representing origin-to-notifier connections as directed networks could help authorities anticipate previously unreported food safety alerts.

The study’s analyses revealed that while directional modeling improved predictions over conventional undirected models, multiplex layering and graph neural networks (GNNs) added negligible predictive value. Standard evaluation schemes also produced misleadingly high performance scores, demonstrating the need for carefully designed comparisons with randomized networks when designing early-warning surveillance systems.

Background

Maintaining food safety across globalized supply chains remains a major international challenge because contaminated products can reach multiple countries before authorities detect a hazard.

Food safety surveillance has traditionally relied on reactive measures (such as identifying contamination through border inspections or investigating its source after foodborne illness outbreaks).

While modern agrifood surveillance systems generate information for proactive decision-making, international supply chains contain complex trade relationships, making it difficult for conventional epidemiological investigations to use these records for predictive modeling.

Previous studies have often treated exporting origin countries and notifying countries as interchangeable or used evaluation protocols that can produce artificially inflated performance scores, leaving the additional value of more complex AI and ML models unverified.

About the study

The researchers sought to determine which components of a network-analytic pipeline genuinely improve the prediction of previously unobserved origin-to-notifier links, thereby informing more targeted international food safety surveillance.

The researchers analyzed 28,724 official notifications recorded in the RASFF database from January 2019 to December 2025. The dataset was structured as a directed network linking countries of origin to countries issuing alerts across 163 countries, excluding domestic notifications from the cross-border link network.

The researchers organized the notifications into 12 hazard-specific network layers (including pesticides, pathogenic microorganisms, mycotoxins, and heavy metals). They then used a rolling-origin link prediction protocol to predict future country pairs, comparing gradient boosting against unsupervised empirical cumulative distribution function (ECDF) rankers and GraphSAGE neural network models.

The analysis compared three negative-sampling protocols (uniform, valid-target, and degree-matched). Model performance was assessed using measures of predictive accuracy and compared with randomized networks that preserved the original patterns of country connections.

Study findings

Accounting for the direction of links between origin and notifying countries improved predictions compared with undirected models. This provided the clearest predictive benefit among the approaches examined.

Neither multiplex layering nor GNN embeddings showed a clear improvement in prediction over the simpler directed-feature model, though the hazard layers remained useful for comparing notification severity.

The study also showed that only 31 of the 163 countries acted as notifiers in the cross-border network. With unconstrained uniform negative sampling, the model appeared to perform exceptionally well. However, randomized networks produced almost identical results, suggesting that the model largely learned which countries could issue alerts and how frequently they appeared in the network, rather than which new country pairs would emerge. Under stricter degree-matched sampling, predictive performance declined substantially, with no clear advantage over the randomized networks.

In an out-of-time evaluation that used data through 2023 to predict newly reported country pairs during 2024–2025, the model correctly identified 45 of its 100 highest-ranked potential links, compared with a baseline occurrence rate of about 9%. These predictions concerned future notification links rather than independently verified contamination incidents. A simple product of exporting out-degree and importing in-degree performed comparably well, identifying 48 of its top 100 predicted links.

Finally, the hazard-layer evaluations revealed marked differences in the proportion of notifications classified as serious or potentially serious, ranging from about 10% for parasites to approximately 95% for mycotoxins and histamine.

Conclusions

The findings show that network topology can help identify previously unreported food safety notification links (at nearly five times the baseline rate), but more complex machine learning architectures provide minimal practical advantage over simpler directional metrics in this dataset.

The authors caution that surveillance databases record notifications rather than contamination events, so raw alert counts reflect regulatory scrutiny and trade volume, neither of which the country-level comparisons accounted for.

Together, these findings indicate that food safety authorities should prioritize inspection resources based on the severity of specific hazards rather than total alert volume. Incorporating external trade-volume and product-level data could help distinguish inspection patterns from underlying contamination risks and inform more reliable international surveillance. The authors also acknowledge that reporting was sparse in 2019 and that hazard classifications were approximate. They caution that the results may not extend to denser trade networks or other prediction tasks.