Food Protection Trends - November/December 2024 - 405

safety plans and systems (37), which poses a major challenge
for the application of AI for food safety. For example, many
firms have not implemented stringent (or any) cleaning and
sanitation breaks between growing lots and reuse substrates
and water without treatments (22). Furthermore, more
automated systems that utilize moving trays with pulleys
and motors (i) are rarely cleaned/sanitized and (ii) are
manufactured in a way that makes equipment challenging
to effectively clean (i.e., filled with hard to reach nooks and
crannies, impossible to reach spots, difficult to take apart and
put together) (37). Lack of separation between growing and
processing, and inadequate food safety knowledge commonly
seen among startups and newer food industries also may
represent a foundational challenge for some facilities. For
facilities at earlier stages of their food safety maturity, it may
not be appropriate to implement high tech data intensive AI
strategies until appropriate foundational food safety practices
are in place, including basic food safety training programs,
which can be highly effective at reducing the risk of microbial
contamination (38). Additionally, creating AI models based
on data gathered from initial CEA production systems
that are lacking food safety foundations may create an
inappropriate starting point for the modelling and AI-based
data analytics efforts, as AI models would need to be fed new
data and potentially reprogrammed to accommodate the
changes that occurred after implementation of foundational
food safety practices. However, even in the early stages of
food safety system development, CEA may benefit from
using existing pre-trained Large Language Models (LLMs),
such as ChatGPT or Gemini, to support basic food safety
tasks, such as personnel training and development of SOPs.
While these models can streamline the tasks, expert reviews
are recommended to authenticate the generated information.
Food safety data availability and quality are important
factors for a well-trained and validated AI tool for food safety
hazards management. In order to acquire the mass of data
needed, a large amount of high-quality data is required to
ensure AI's reliability; the only efficient pathway to this may
often include data sharing between firms. This poses the
challenge of data privacy hesitations. Food safety data are
highly sensitive, due to fears of data abuse, bad publicity,
reputation, liability, and the need to keep certain data (e.g.,
human illness data) confidential (1, 53). CEA and other
food companies can increasingly recognize the value of
data sharing based on successes in other industries (from
medicine to hospitality industries), where sharing of data
has allowed scaling of AI applications and learning through
peer networks (53). Improved data sharing, including the use
of shared data in AI models, has the potential to (i) provide
food safety benchmarks for the industry and (ii) facilitate
better business and food safety decisions. Thus, there is
a need for research geared toward a better understanding
of data sharing obstacles and the development of data
infrastructure and algorithms that secure the privacy of
users who engage in data sharing. One way of addressing
this challenge is through Federated Learning (FL), which
has gained attention in several domains (8, 19, 21, 31, 70).
In a federated environment, data remains secure within the
physical location (i.e., data station) of its owners. Instead of
transferring data, the model moves between these locations,
effectively updating the model parameters from the data at
the respective data stations, abiding by privacy principles.
For example, Gavai et al. (31) developed a Federated
Bayesian Network (BN) model to predict food fraud, which
demonstrated the applicability of the federated BN in food
fraud; they anticipated that such a framework may support
stakeholders in the food supply chain for better decisionmaking
regarding food safety control while still preserving
the privacy and confidentiality nature of these data. In
addition to the amount of data needed, the quality of data
is also imperative. To preserve data quality, monitoring data
needs to be consistent and standardized throughout the data
collection period and even beyond the project time to enable
a sustainable data source for further model improvement and
validation. This can be challenging to apply from firm to firm
as each company will have different data collection methods,
data labeling, and different streams of data. This limits AI's
applicability to the sector. While there is a growing trend
of high-tech precision farms, there still remains a number
of firms that use traditional produce growing and data
monitoring methods such as pen and paper data collection.
Some firms simply lack monitoring of growing variables
altogether. A unique aspect of food safety data, especially
microbial concentration data, is that contamination, while
serious, is often rare. Contamination with pathogens would
be typically detected in only a few samples, and the majority
of the samples would be non-contaminated or contaminated
below detection limits. These unbalanced datasets need to be
handled carefully in the modelling process to reach a desired
prediction accuracy for the positive samples.
It is important to realize the need to develop human
resources in parallel with the development of technologies
for improving the food safety of CEA-grown produce. This
emphasizes the need for cross-disciplinary training between
domain knowledge and data science/engineering disciplines.
Additionally, to be able to leverage the full potential of AI
technologies in CEA food safety, these technologies will need
to present users with interpretable and useful information
through an effective human-machine interface. Not only
will users need to interpret results, but they will need to
understand the models to address concerns about false
positives and business liability as well.
Finally, to drive appropriate adoption of AI, economic
aspects and costs and benefits of adopting AI systems may
need to be quantified, including to understand opportunity
costs that may be associated with implementing AI to
help assure food safety (as there may be other food safety
investments that generate a greater risk reduction in return
November/December Food Protection Trends 405
Food Protection Trends - November/December 2024

Table of Contents for the Digital Edition of Food Protection Trends - November/December 2024
