Using EU Rapid Alert System for Food and Feed (RASFF) data, researchers have demonstrated the utility of artificial intelligence (AI) technology for food safety risk assessments. Published in Food and Bioprocess Technology, the study was authored by researchers from the University of Portsmouth and the Institute for Food Science and Technology (IFST).

With an increasing amount of food safety data available, AI techniques are being explored as an important solution for processing and analyzing these large pools of data to improve risk assessment and decision-making, ultimately enabling more informed food safety management and improving overall food safety. However, more research is required to understand the differences in the effectiveness of different AI technologies—like machine learning, deep learning, and explainable AI (XAI)—as well as to overcome the existing limitations to their predictive accuracy.

In this context, the present study aimed to develop an advanced food safety risk prediction system to improve traditional food safety risk assessment methods and decrease the ambiguity of risk management decisions. Using the RASFF dataset, the researchers explored the integration of machine learning, deep learning, and transformer-based models, enhanced with XAI techniques, to improve both model transparency and interpretability.

Established in 1979, RASFF facilitates the exchange of information between EU Member States regarding food safety incidents in support of swift response by authorities to mitigate risks to public health. Containing more than 61,000 entries, the database is a comprehensive repository of notifications regarding food and feed safety concerns, including risks identified both within the EU and risks originating from outside the EU.

Overall, the researchers were able to demonstrate the potential of combining advanced machine learning, deep learning, and transformer-based models with XAI techniques for food safety risk assessment. The critical importance of data enrichment, advanced architectures, and model interpretability in improving prediction accuracy and fostering trust in automated systems was noted.

Data enrichment was especially pivotal to improving model performance across all categories. By addressing challenges such as short explanations and underrepresented classes in the RASFF dataset, data augmentation contributed to enhanced generalizability and classification accuracy. Importantly, enriched datasets not only improved predictive capabilities, but also uncovered additional critical features such as heavy metals and botulinum toxin that strengthened the model’s decision-making process.

Moreover, XAI techniques (e.g., SHAP) helped elucidate the models’ decision-making processes, improving transparency and interpretability. By identifying the most influential features for risk assessment decisions, XAI provided actionable insights that are highly relevant to food safety regulatory bodies; for example, features like Salmonella, aflatoxins, and Listeria monocytogenes were associated with serious risk management decisions.

Comparatively, transformer-based models (e.g., BERT and RoBERTa) outperformed both traditional machine learning models and deep learning architectures, and their accuracy and robustness in handling complex and highly dimensional textual data make them most suitable for nuanced tasks like food safety classification. On the other hand, traditional machine learning models, while still effective, had greater limitations when presented with imbalanced and intricate datasets.

Despite the promise of the integrated AI approach to food safety risk assessment demonstrated in the study, roadblocks to the practical implementation of such models still exist. Specifically, the computational complexity of models like BERT and RoBERTa require substantial resources and may be impractical in resource-constrained settings. Additionally, adoption of these techniques by regulators who prefer more transparent models may be a challenge, underlining the importance of explainable AI models. Scalability is also a technical concern due to different data formats, standards, and legal frameworks used in different regions. Finally, the potential for misleading outcomes arising from training data biases is always an issue, and generalizing the approach to new or significantly different datasets may require additional training or fine-tuning due to regional divergences.

The researchers recommend further exploration to improve their AI framework for food safety risk assessment, such as expanding the dataset to include additional contextual factors like geographic data, supply chain information, and temporal trends. Another area for future work is the application of these methods to real-time data streams, enabling continuous monitoring and assessment of food safety risks.