Predictive risk assessment of food using artificial intelligence-based data models

Annually, approximately 600 million cases of foodborne diseases are reported worldwide, with up to 420,000 resulting in mortality. Reducing these numbers and enhancing food safety remain a global concern. In the current era of digitalization and technological progress, new opportunities have emerged for being applicable in enhancing the safety of food. The volume of digitalized and free available data relevant to food safety is continuously increasing, offering valuable resources for risk assessment and mitigation. However, these large amounts of data also present a challenge, as the manual analysis of such vast quantities of data can necessitates major amounts of time and resources. This is where the use of artificial intelligence (AI) can represent a great opportunity. The objective of this thesis, along with the studies carried out, was to investigate whether, and if so, to which extent, an AI-supported database can contribute to the early detection of food safety risks. Initially, the capabilities of the AI-supported database that was mainly employed in this study were evaluated along with a presentation of the existing approaches for early risk detection. Exemplarily using the limit value exceedances of ethylene oxide (ETO) in sesame seeds during the year 2020, the suitability of the database for detecting this issue before 2020 was assessed. Using AI-generated data feeds, initial indications of ETO limit value exceedances would have been available as early as 2018, suggesting that these data feeds could have provided early warnings much earlier. Even though these initial indications emerged in black pepper powder, they could represent an early indication of a potential ETO issue in sesame seeds, as fumigants like ETO are not exclusively applied to a single food product or matrix. However, a substantial number of new incidence reports are published daily, and manually sorting these reports presents a major challenge in terms of time. Nevertheless, to identify food risks as early as possible, it is crucial to evaluate the content of these reports. Nonetheless, not every incidence report describes an immediate food risk or necessitates urgent action to mitigate potential harm to businesses and consumers. To differentiate incidence reports with hazardous food safety risks from less critical ones and eliminate the need for manual screening, a risk assessment framework for incidence reports was developed. After an initial testing phase, this framework was aimed to be integrated into an AI-supported database to enable time-efficient, automated evaluation. As part of the manual test run, various incidence reports for the commodity sesame seeds were exemplarily assessed and categorized based on their potential risk. The automated assessment optimizes resource allocation by reducing time demands. To further expand this benefit, the risk assessment framework was enhanced with additional indicators, enabling its use for identifying reliable suppliers within the food industry. Collaboration with unreliable suppliers – whether due to their own actions or external factors – can result in the distribution of food unsafe for consumption, thereby compromising food safety. To prevent the introduction of unsafe food products into the market, it is essential to establish partnerships exclusively with reliable and verified suppliers. However, the manual process of searching for and verifying suppliers is time-consuming. Therefore, the AI-supported risk assessment represents a major advancement, enabling more efficient and time-saving identification of reliable suppliers. To implement this assessment scheme, the evaluation framework for incidence reports was expanded with six additional indicators, which are suitable for providing insights into the reliability of a supplier. The indicators include the incident category level, indicating how widespread the goods are already on the market, the vulnerability of a commodity of interest, the potential audit performance of the supplier, the logistic performance index (LPI), the gross domestic product (GDP) growth, and the GDP per capita. Each indicator is assessed independently and incorporated into the calculation of the overall risk, with an individual weighting. Initially, the supplier risk assessment scheme was tested manually using exemplarily selected suppliers. This approach facilitated the appropriate categorization of suppliers’ risks, ensuring its suitability for subsequent integration into the AI-supported database. The approaches developed provide suitable methods for an AI-supported earlier detection of food safety risks. However, it is important to note that these methods are only suitable for recognizing early warnings and to some extent for upcoming or novel risks. Particularly in the case of early warnings, the food safety incidents have typically already occurred. Initial individuals may have already been affected, necessitating measures to mitigate the further spread and impact of the damage. However, to enhance food safety and reduce the incidence of foodborne diseases, it is crucial to anticipate and predict future risks that remain unknown. Different approaches can be considered for leveraging an AI-assisted database in the recognition of foresight risks. Of particular importance will be the extraction and analysis of both scientific datasets and, where applicable, information derived from social media sources. Achieving this capability will necessitate the future development of advanced algorithms and procedural frameworks to enable the AI-supported database to function autonomously in detecting potential foresight risks.