The recent FDA podcast on “Artificial Intelligence in the New Era of Smarter Food Safety” highlights several key benefits of sharing data in the realm of food safety:

Enhanced Predictive Analytics: Sharing data allows for the development of more robust AI and machine learning models. These models can more accurately predict food safety risks, helping in the early detection and prevention of potential foodborne illnesses or contamination.

Improved Horizon Scanning: By pooling data from various sources, organizations can perform more effective horizon scanning. This process identifies emerging risks in the food supply chain, enabling proactive measures for risk mitigation. Shared data can offer a more comprehensive view of potential issues, leading to better-informed decisions.

Collaborative Problem-Solving: Data sharing facilitates collaboration between different stakeholders in the food industry. By working together and sharing insights, companies, regulatory bodies, and other organizations can collectively address food safety challenges more effectively than they could individually.

Resource Optimization: Access to a larger pool of shared data allows organizations to optimize their resources. They can focus their efforts where the data indicates the highest risk, leading to more efficient use of time and financial resources in ensuring food safety.

Benchmarking and Best Practices: Shared data enables organizations to benchmark their practices against industry standards and learn from the best practices of others. This can drive continuous improvement in food safety processes across the industry.

Democratization of Food Safety: The podcast emphasizes the concept of democratizing big data in food safety. This approach makes valuable data accessible to a wider range of stakeholders, including smaller companies that may not have the resources to collect extensive data independently.

Global Perspective: Data sharing, especially on an international scale, provides a global perspective on food safety challenges. This is crucial in a world where food supply chains are increasingly globalized, and risks in one part of the world can quickly become international issues.

Overall, the podcast underscores the immense value of data sharing in advancing food safety, promoting innovation, and fostering a more collaborative and proactive approach to managing food safety risks.

The speakers:

• Frank Yiannas, Deputy FDA Commissioner for Food Policy and Response, and Donald Prater, Associate Commissioner for Imported Food Safety, led a discussion with food industry experts on subjects that include the opportunities that AI offers to help protect consumers from food safety issues, potential uses of AI that food producers could consider, and what’s on the horizon for AI in FDA’s New Era of Smarter Food Safety.
• Maria Velissariou, Global Corporate Research & Development Vice President and Chief Science Officer for Mars Incorporated, a global, family-owned business with a portfolio of confectionary, food and pet-care products and services;
• Nikos Manouselis, founder and CEO of Agroknow, a food safety intelligence company that predicts food safety risks to inform preventive measures; and
• Cronan McNamara – founder and CEO of Creme Global, a company providing food safety data analytics and predictive modeling software and services. 

Automatic transcript:

Artificial Intelligence in the New Era of Smarter Food Safety

[00:00:08] Frank Yiannas: I’m Frank Yiannas, Deputy Commissioner for the Office of Food Policy and Response. And I’m here with Dr. Don Prater, Associate Commissioner for Import and Food Safety at FDA. We’ll be co hosting today’s episode, which will focus on artificial intelligence and how it can contribute to the safety of the food we all eat.

I think any of you that have heard me speak publicly about AI and food safety during my tenure here at the agency will know that this is a topic I’m pretty passionate about. Why? Well, I see the use of AI as an absolute game changer. Powerful new tool that we can add to our food safety toolbox, one that could significantly enhance our ability to create a safer food system.

And AI is actually an important element in our goals that we’ve set forth under the new era of Smarter Food Safety, something that we’re doing to bend the curve of foodborne illness once and for all in this country and around the world.

When we look at how other industries are harnessing the power of data to identify and predict trends, it’s clear to me, as I suspect it will be you, that the FDA and food system stakeholders should also be looking at how to tap into new technologies such as artificial intelligence. 

At FDA, we’re interested in strengthening our predictive analytic capabilities through expanded use of AI and machine learning tools, and to use AI to mine information from non-traditional sources such as social media and apps, detect outbreaks and to supplement traditional health reporting. 

In the next hour, we’re going to explore along with our guests opportunities that AI offers to protect consumers from contaminated foods, potential uses of AI that food producers can start considering right now, not tomorrow. And what’s on the horizon for AI in both government and industry.

And so, with that, let’s get started. Don Prater, over to you.

[00:02:04] Don Prater: Thanks, Frank. Well, as Frank mentioned, I’m FDA’s Associate Commissioner for Imported Food Safety. And I help oversee that portion of the U.S. food supply that’s imported from abroad, for which the percentage of FDA regulated food products stands at about 15% and growing.

However, for certain commodities, those percentages are really outsized. In 2019, those shipments included 32% of our fresh vegetables, 55% of our fresh fruit, and more than 90% of the seafood Americans love to consume. Later in this podcast, I’ll talk more about some areas that FDA is exploring with respect to import screening using AI and machine learning.

And now I have the privilege to introduce our distinguished panel of experts who will share their unique experiences and insights on AI and food safety. 

Our guests today are Maria Velissariou, the Global Corporate Research and Development Vice President and Chief Science Officer for Mars, Nikos Manouselis, Founder and CEO of Agrino, a food safety intelligence company that extracts tailor made data insights for the global supply chain, and Cronan McNamara, founder and CEO of Creme Global, a company providing food safety data analytics and predictive modeling software and services. 

So, thank you all very much for joining us. Let me begin by asking each of you to share with us ways in which industry is pioneering the use of AI to protect consumers from unsafe food. 

Pioneering AI for Safer Food: How Industry Leads the Charge

And would you please also provide a specific example in which you’ve seen AI. Make a real difference. Maria, would you like to start us off?  

[00:03:44] Maria Velissariou: Thank you, Frank and Don. I’m honored to be here today, and it is a pleasure to be with Cronan and Nikos, too. AI Is becoming increasingly embedded in the entrance supply chains and agriculture and food.

It gives us algorithms that when combined with conventional techniques like forecasting can sharpen and expedite foresights and insights. And AI powers the internet of things can improve efficiencies detect defective or unsafe ingredients in food processing. And ensure that food safety protocols are adhered to in compliance with regulations.

The technology has a lot of promise for food safety, but we need to ensure that we have high quality, accurate, and secure data. And that we have addressed factors like human biases. Leveraging AI allows companies like Mars to make food safer. Available to more people and in a more sustainable way by helping reduce the environmental footprint and waste. 

Horizon Scanning for insights and risk analysis

An example that I would like to highlight is AI horizon scanning that generates foresights and insights for food safety and risk mitigation. The value of horizon scanning comes from the data generated insights, the indepth risk analysis, and the action that it helps drive. And also horizon scanning can provide more transparency and holistic perspective on recalls and withdrawals and in this way facilitate learning and improvement within an organization and across the industry. 

Aflatoxin Prediction

I would also like to highlight another example, which is aflatoxin prediction. Currently, we have an aflatoxin predictive model at Mars, which we have deployed within Supply Quality Assurance to guide us with sourcing options. We’re in the process of using AI based algorithms to enable us to see patterns and move into predictive analytics.

The model is based on high resolution meteorological, geospatial, and temporal data. It predicts aflatoxin generation in the field, during transportation, and in storage. Here, the ultimate goal is to provide farmers with preventative tools to mitigate formation in the first place. And to this end, we are exploring come to bid pilots with our Food Safety Coalition partners, which involves Sacatinia, a nonprofit organization. 

Over to you Don. 

[00:06:14] Don Prater: Thanks Maria. Those are great examples. I really appreciate hearing about how Mars is using horizon scanning and in particular, the aflatoxin prediction model. Nikos, what’s your experience been?

[00:06:27] Nikos Manouselis: Don and Frank, thank you so much for the kind invite. It’s a pleasure and honor to join this panel.

Horizon Scanning – Web Crawling Software 

My example comes also from the area of horizon scanning risk horizon scanning, but I want to dive deeper into, the types of incidents that are being announced all over the world, not the ones officially announced by the major agencies or, authorities like the FDA or the European [inaudible] portal, but looking at these small authorities, the announcements that come from very local municipalities that often are being published in some website in one of the national languages or dialects, and they can be very, unstructured so much. So, how can we use AI in order to discover and get this type of information also in horizon scanning software? This is where we use web crawling software systems that employ AI in different steps of such a process.

They can identify using NLP technologies (Natural Language Processing Technologies), what exactly is being described in an announcement. They can translate an announcement that comes in a different language into English. They can use techniques to annotate what is the particular product? What is the company that has been involved? What is the hazard that we are referring to? So that they create very structured data records. They can also combine announcements that come from different sources on the same incident so that they can create a more rich, enhanced description of something that is important for someone to know. 

By employing these techniques, we can have better information and wider representation in near real time compared to the hours or sometimes days that a human would require in order to perform such a task. And I already can think of a couple of examples of clients, like a very large beverage manufacturer that is using this real time intelligence to inform their quality council, the experts meeting in the organization that needs this external risk intelligence on a periodic basis so that they can take decisions. Or another food service chain that is using this information and incorporates it in a monthly newsletter that circulates to all the food safety professionals in its ecosystem. 

[00:09:30] Don Prater: Thanks, Nikos. That’s a great example, and I really appreciate hearing more about how you’re looking at the examination of food safety incidents, not only at the national level, but also those that are reported by the local level. 

We’re also interested in natural language processing, one of the disciplines of AI, and also the power of AI to look at structured and unstructured data. So thank you very much for sharing that.

Cronan, over to you. What about your thoughts?

[00:09:55] Cronan McNamara: Thanks, Don and thanks, Frank, for the invitation to be here. Yeah, our experience in this field is, well, Creme Global has been working with industry for many years and governments to help gather anonymized and structure data so that it can be shared and visualized and in recent years, we’ve been aggregating these industry data sets in order to train up machine learning and AI models on the data so that we can, you know, uncover hidden patterns and make predictions. 

Western Growers Food Safety Data Sharing Project

So I have two examples to share with you today. My first example is the Western Growers Food Safety Data Sharing Project. This will be a project that you’re well familiar with, Frank. The FDA did provide letter support for the project stating that it is in line with your vision for the future of food safety. 

So it’s a very exciting project that Western Growers Group is organizing to collect data. And we started with a pilot program of leafy green growers in the California and Arizona region. And now that we’ve aggregated a number of growers data, we’ve created dashboards where these growers can come in and visualize what’s going on with their data, at an aggregated and anonymized levels. 

The types of data we collected are information on things like inspections, product testing, water testing, and location, and these are now being combined with information from weather, topology, dates, and seasons.

And now we’re training up a machine learning tool to uncover risks and trends that may not be apparent to the human eye, even when you’re visualizing these data sets. So we’re able to now predict risks that are emerging or increasing and allow the growers to understand and benchmark their operations against their peers and their colleagues in the region and understand emerging risks in that region.

Sequence Alliance for Food Environments (SAFE)

The second example is a project called the Sequence Alliance for Food Environments, our SAFE project here in Ireland, which was a collaboration between ourselves, Creme Global and UCD, and six leading food companies. And this was more focused on the factory environment. And the goal of the project was to develop a predictive software toolbox to enhance food safety and quality using environmental information, and in particular using next generation sequencing.

So, we used, 16SRNA methods to swab and gather data at the genus level. And because that data is quite complex, combining that then with other information like IOT sensors and traditional microbiological culture results, we were able to use machine learning again in this case to really understand the microbiome of those environments. And to make predictions of when that microbiome may be changing the emerging risks of more dangerous pathogens or bacteria evolving in those microbiomes. So in short, the project helps the industry to understand their manufacturing environment in more detail on the machine learning and data management, data analytics methods help them to really understand the potential and emerging risks in those environments and to take mitigating actions in good time.

So those are 2 key examples that I’d like to share today. Thanks.

[00:13:31] Frank Yiannas: Thanks Cronan and Maria and Nikos. Those are really great and compelling examples. Really, really energizing. 

FDA and AI – Seafood Pilot

We’ve also received some questions from stakeholders about how FDA in particular is using AI. So before we move on, Don, I was wondering if you could tell our listeners. What are we doing with AI and in particular around the seafood pilot?

[00:13:53] Don Prater: Sure. Thanks, Frank. FDA’s been conducting a pilot that leverages AI and specifically machine learning to strengthen our ability to predict which shipments of imported food pose the greatest risk of violation. In August 2020, FDA shared some initial findings from a proof of concept for imported seafood in which we trained a machine learning model against two years of data associated with seafood imports and then tested it against the subsequent year’s data.

The proof of concept suggested that machine learning could greatly increase the likelihood of FDA finding a shipment containing potentially contaminated products. We started with seafood because, as I mentioned, in the United States, we import a significant proportion of our seafood. In fact, over a million lines of seafood are offered each year, which translates into thousands of entry lines each day, for which FDA must make admissibility decisions, including which shipments to examine, sample, and test.

From February through July 2021, FDA conducted a second phase of the AI imported seafood pilot in the field. We wanted to further examine the deployment of an AI machine learning model, and importantly, its integration into FDA systems and operations. We’re currently reviewing these findings and may have more to say about that at a later stage.

But more broadly, FDA is looking to apply lessons learned from the seafood pilot to enhance our predictive capability for other types of foods and regulated products. So stay tuned to this space. Back over to you, Frank. 

[00:15:27] Frank Yiannas: Hey, thanks, Don. I’m very proud of the work the FDA is doing using AI as part of FDA’s New Era of Smarter Food Safety.

To everyone listening, I just think what you’ve heard underscores the promise of AI and how it can truly, truly help to strengthen food safety. Now, some of you might be wondering, as am I, that sounds great, but I know there are people representing a variety of different stakeholders, industry, food production facilities, and you might be saying, well, how can I use AI for what I’m doing in my workplace?

How to Use AI at Work 

And so let’s explore that for just a bit. Nikos, we understand that AI is wonderful, but it’s more than just the technology. It’s not that the listeners are chasing the new trend or shiny coin. They want to know how to use this to solve some of their biggest public health challenges. And so what type of, use cases do you think AI can help them address? And once they’ve identified the use case, do you have any specific, recommendations for listeners on how to get started? 

Emerging Risk Prevention

[00:16:30] Nikos Manouselis: Great use case, or public health use case, has to do with emerging risk prevention, at least rapid mitigation, but ideally prevention. So, how can we put in place services powered by AI, facilitated by AI that will help us respond quickly or take preventive measures before something hits consumers and the public. 

And my favorite analogy is weather forecasting and the way that a combination of public sector services and infrastructural components and pieces, and private sector services and, infrastructures are putting in place together so that they can help us, for foresee forecast and address natural disasters. 

So how could we put in place a similar array of infrastructural layers so that we can do something, as important as food risk prevention. I would invite our listeners, regardless of whether they work in the public sector or in the private sector to start decomposing using such an analogy, the layers of infrastructure that we need. 

Where does data reside? How can we collect and get together in trusted data resources, in databases or in larger structures? All this information that is important for us so that we can feed predictive models. And then how can we develop the right models for its purpose? How can we combine scientific, public and private forces and, intellectual power so that we can develop those predictive models that will help us calculate the likelihood of an event that is coming up in the supply chain.

And at the end of the day, how do these pieces come together so that we can power an ecosystem of services? Public services, privately developed and monetary services so that we can use these services in given business problems in specific decision scenarios to address and try to help answer specific questions. 

So, I will not think about a given solution to a specific problem. I’m trying to think about use cases that are solving very grand challenges, by decomposing them in the different components that have to come together. 

[00:19:40] Frank Yiannas: Thank you, Nikos. That’s wonderful. We often say food safety requires collaboration. And you’re just saying, hey, we might be able to collaborate on the use of this new and powerful technology. 

Maria, let’s start, you and I have talked about AI in the past, and I was struck in our conversations of some points you emphasize with me that, you know, to be successful in an IT project, it’s more than just the IT. It also involves the human element. And so, you persuaded me that this is something we should talk about.

So, who else in the organization, if people are listening how we get started, who else should they be including, to consider some of these union or behavioral aspects of a transition to AI? 

Partnership Between Several Functions 

[00:20:23] Maria Velissariou: Thank you for the question, Frank. Secondly, this is an interdisciplinary area that requires a lot of collaboration.

At Mars, we believe that data and AI and different functions coming together are absolutely vital components for tackling supply chain challenges, both tactical and strategic. And we want to empower our associates to move 100 times faster today, which we cannot do without the power of data. And the power off analytics. 

Artificial intelligence generates a lot of value and adopting the approach to maximize this value requires the partnership between several functions like R & D, digital technologies, manufacturing, procurement, sales, among others. 

As Nikos and Cronan highlighted earlier, there is a plethora of data out there. And it is important to detect signal from noise and establish the right actions in a timely fashion. Today, we have incorporated artificial intelligence in more than 250 projects and capabilities across Mars, ranging from healthcare, sustainability, consumer ecosystems, to name but a few. It is indeed a key enabler to unlock great speed and value.

We also need to have a willingness to experiment quickly [inaudible] what doesn’t work and scale fast what does. And we need to push through the hype to understand what problem we’re really solving. Not every problem requires an AI solution, and that’s why it is critical in the process to employ a user centric approach.

There may be also a couple of other areas that we need to consider, and upscaling is one of them. I associate with deep expertise in traditional sciences like microbiology and toxicology may need to become fluent in the basics of AI so that they can understand how AI generates value. They don’t need to become AI engineers.

Equally, AI engineers need to understand the basics of food science and technology. So that they can collaborate with their counterparts. And importantly, food safety is a great example where AI can assist human judgment, but cannot replace it. Nor is AI a substitute for good behaviors and practices, which are cornerstones of culture.

These need to be practiced and modeled every day and consistently.

[00:22:54] Frank Yiannas: That’s excellent. Thank you, Maria. 

All right, Cronan moving over to you. You and I have talked about data for many years now. You’ve given me great advice and based on our conversations, I’ve coined this term, that I’ll attribute to you, which is AI sometimes seems magical, but it doesn’t magically happen.

And you’ve persuaded me that the quality of data is foundational and critical. So I was hoping that, I’ve heard it before, but I was hoping that you could share with our audience, uh, you know, any tips on why this issue of data quality is so important. 

Data Quality is Important 

[00:23:28] Cronan McNamara: For sure, Frank. And that’s so true. And you asked me to keep it concrete.

So I’m going to try and give you some good concrete tips on on data quality and how companies and organizations can manage that. 

Carry Out Comprehensive Checks on Your Data

So, of course, first of all, it’s important to carry out really comprehensive checks on your data. We’ve all heard the same garbage in garbage out. We don’t want the data to lead our models astray.

And this can be especially risky if there are many steps in the data collection process. But so how do you do that? Well, let’s start with trying to have a consistent format for your data sources, and this helps to, um, match up data in a supply chain and also helps error checking and you should review outlier data and question is, is it a data entry error or is it a valid data point that’s an outlier in your data. Because it’s an important distinction. 

Review all of your missing data and no data values. What did these really mean? Is it a negative result? Or is it just missing data where no test was performed, for example. And data engineering now has become an engineering process where you have to build in quality assurance checks into your data engineering process.

So a good way to check your input data is using visualization. So visualizing your input data. It can be quite important as important as visualizing the results, because when you visualize it, you start to see the patterns and you can check quickly because the data makes sense and you have to make sure you really understand the meaning of each data point and what assumptions are implicit in the data.

So ideas to help with that. You know, read the documentation on the data or even talk to the person or the team who collected the data. All of the metadata that comes with the data is also so valuable in the model. So make sure you understand those variables as well. 

The issue as well, even though we have a lot of data these days is bias in the data and instead of garbage in garbage out, I like to think of bias in, bias out. If the input data is biased, the results of your model can be biased.

So often the data you’ll be using in your model was collected for a different purpose. It could have been collected for a marketing purpose or a monitoring program purpose and you’re trying to use it for food safety. So checking things for bias, you know, does it cover all scenarios? Does it cover all seasons or was it just done during the summer, for example, are important things to check.

Finally, when there are rare events that you’re interested in your data, you need to have a large amount of data. So quantity is also important as well as the quality of the data so that you have enough samples of those rare events. So those are my, I suppose, top tips, Frank, for data quality in trying to build machine learning and AI models.

Future of AI

[00:26:24] Don Prater: So now let’s focus on the future of AI. Maria, what’s on the horizon? How is this field [inaudible]?

[00:26:33] Maria Velissariou: Thank you, Don. At Mars we’re guided by our five principles with quality being the first principle, and we want to use AI to fear associates from low value, repetitive tasks so that they can focus on higher value, creative work and in this way, move into deeper analysis and judgment.

This means that we can use AI to organize huge amounts of data coming from, raw materials and manufacturing processes, so food safety experts can apply their domain specific expertise to interpret the data. It’s important to identify bias where it is and sort it out early on in the process.

Digital Quality and Food Safety – Mission Control

I would like to use three examples to bring this to life. The first example is about digital quality and food safety. We’re leveraging the power of data to ensure high quality and safe food production through a platform called Mission Control. The platform analyzes critical performance metrics, performs comparative analysis, and produces new insights in a way that was not previously possible.

And we are focusing on four categories of exploration: 

• Raw materials like the aflatoxin predictive model that I talked about earlier on
• Quality Analytics 
• Finished Products 
• And Horizon Scanning and External Listening. 

And we’re currently driving end user adoption and working on machine learning for the next generation of capabilities, like risk profile of raw materials.

Traceability 

The second example is coming from the area of traceability. We have deployed systems and processes, several of them, to enable traceability of raw materials and finished products. However, we’re still relying heavily on human resources to execute the process against standard operating procedures, and we know that the vast majority of events are caused by non compliance to procedures.

By using process mining, we can see how the process has been practiced on the shop floor, and therefore put in place corrective actions to increase compliance and in this way, uh, increase business performance. 

Culture

My third example is about culture. We want to live by example as we drive our digital transformation. And we asked our senior leaders to sponsor the change coming from digitalization, like predictive and prescriptive analytics. And we have decided to implement reverse mentoring to give the senior leaders the opportunity to be coached by internal digital experts. And this is an effective way to help them play their role as sponsors in this digital transformation journey. 

The second aspect of culture is an inaugural Mars Artificial Intelligence Festival that we ran in 2020. It was a week long, all virtual immersive experience to educate, demystify, and celebrate AI among all Mars associates. It was actually a great opportunity to showcase how prevalent AI already is within the organization.

And there is also a lot of learning from each other. And I want to quote here two examples from two very different parts of the organization. One is from our Mars Veterinary Hospital. In the anatomic pathology division, we have created AI use cases for tumor detection. And now we can read x-rays in a matter of minutes as opposed to days.

The second example comes from the other side of the business, from product manufacturing, where we use AI to measure the color blend of schedules, ensuring consistent quality. Once again, these allow our associates to focus on more strategic work. 

[00:30:39] Don Prater: Thanks, Maria. I think those are really terrific examples. Very specific. And I know they’ll be helpful to our listeners. 

Insights into the Supply Chain 

So, Nikos, can you provide some insights into the supply chain? How is more visibility in times of pandemic, severe weather, also being able to be impacted by our new technologies such as AI and machine learning? 

[00:31:03] Nikos Manouselis: Those that can be more systemic and holistic in their view. So, how can we look at the supply chain as a system that is affected and influenced by different factors that then lead to an emerging risk? 

Example: Major Beef Producer

I can share two examples there. One is the example of a major beef producer in the UK that was severely hit by the horse meat scandal. And they took the decision to finance scientific research that will help them understand and map the different paths in the supply chain that had to do with red meat and then map out the signals that could influence fraud and maybe serve as predictors of fraud as far as beef is concerned. 

And when they did this exercise, then they came to us and they said, okay, how can we use these candidate signals in a model that will be behind a dashboard that will operate in real time. And we did a quite elaborate model development exercise where we started with the obvious signals, like the prices of beef in the different markets from where they were supplying. 

But then we started looking into other systemic signals, like, did we have any political events that had an effect on prices? Did we have any other type of indicators like country corruption risk that we could associate with the prices of beef? And through this exercise, we developed a model that then was proven to be able to predict fraud events in a way that was very, very specific to their needs. 

Example: Wageningen University in the Netherlands

I have another example in mind. This is a model that has been developed by Wageningen University in the Netherlands that is trying to look three to five years down the road and see, can we forecast the levels of residues that we will see in food if we start by other type of earlier indicators, like the level of consumption of, chemical products in a given market.

And they did this again, modeling exercise with lots of data about the Netherlands. And they saw that they could develop, and they did develop a model that is taking as input different parameters, like how many products were consumed at the production level, agrochemical products were consumed at the production level. And this served as a forecaster, as an indicator of that helped them see the level of residues that they would expect a few years down the road. 

[00:34:20] Don Prater: Thanks, Nikos. Those are some really interesting examples. 

SME’s and AI

Cronan, can you tell us about small and medium sized companies? Can they use AI? What’s required in terms of technology and expertise? And does open source data make this a more feasible option? 

[00:34:37] Cronan McNamara: Sure, Don. And absolutely, yes, small and medium sized companies are already starting to benefit from AI, even if they’re not necessarily building it in-house or haven’t gotten currently have the expertise or capacity in-house. 

So there are a number of industry wide projects that have taken place that are providing and developing AI models that members of an industry group can then tap into and benefit from. So these are kind of semi-public and industrial systems that they can access. And, of course, there are companies like Agrino and ourselves, Creme Global, that specialize in food safety and have the expertise and infrastructure to help companies get started on the mission to start using their own data with some models that have been developed for industry are more widely developed models. 

Build Your Own In-House Team

And they can, of course, then start working towards developing their own in-house team. Like Maria has in her team a very well established machine learning and AI team. 

But if you want to start from trying to build up a team, I think some of the points that were made earlier are really important. You know, you would need to assign a fairly good scientist to be responsible for this team who probably understands the food safety and the microbiology or toxicology of the situation. And then start to recruit and train up, the following roles. 

Data engineering. I’d start with someone who can start to organize all the data in the organization. Next, you would want somebody with really good grounding in mathematics or physics who can take on the AI piece. And you would need some software visualization expertise and then some computing system experts to manage all of this. 

Could you find all of that in one person? It’s possible, but very hard. It’s very rare to find. So probably you’re talking about hiring 2 or 3 people there. And so it’s not a trivial undertaking and that would take some investment. 

Open Source Data

Open source data is a really interesting way to start and there is very valuable open source data being provided in the US by organizations such as the CDC, yourselves in the FDA, USDA. And what’s great about the open source data is it provides a starting point and direct access to well structured data that can kickstart the data repository for these companies.

And then they can start trying to curate and aggregate their own data and combine it with this open source data in order to provide an even richer data source that’s more relevant to their company’s operations. And open source data is very powerful when it’s overlaid and combined with industry data or company data.

So these industry groups who get together and start aggregating and analyzing data and combining with open source data, I think are using all of the data resources that they can get access to, to create predictive models that can benefit their whole sector and the companies within us. 

[00:37:42] Don Prater: Thanks, Cronan. That’s really fantastic. It’s good to hear that there are opportunities for small and medium sized companies to get engaged and utilize artificial intelligence in this way. 

So I’d like to thank all of our panelists for today. Really appreciate sharing these great examples. I think it’ll be super helpful to our listeners.

And so I’d like to turn back to Frank now for some of your thoughts on today’s tech talk. Frank?  

Conclusion

[00:38:11] Frank Yiannas: All right. Well, thanks so much, Don, Maria, Nikos, and Cronan. I think it’s clear from today’s podcast that I’m not the only one who’s passionate about AI and its potential to tackle food safety challenges. All I can say is, wow, there was so much discussed today.

I’m super energized by the conversation, but I thought I’d close by just sharing with the audience some of my key takeaways, and I’ve distilled it down to five, literally I’ve been sitting here taking notes as you guys were speaking. 

Better Food Safety Begins and Ends with Better Data

Number one, better food safety begins and ends with better data is what I heard. Better food safety begins and ends with better data. I know we’ve used great tools in the past, inspection approaches, the tool of training, the tool of testing, but we’re entering the 21st century where we increasingly have the ability to convert large volumes of data into powerful, predictive information through tools such as AI. And so we emphasize the quality of data. So better food safety in the 21st century is going to begin with better data. 

The Public Health Problem We Are Trying to Solve Using AI

The second key takeaway for me was that, well, while we’re talking about AI as a technology, what you’ve persuaded me, it’s not about the technology, it’s about the public health problem that we’re trying to solve. Not just about technology, it’s the public health problem that we’re trying to solve. And I think you guys did a wonderful job of giving some concrete examples of specific use cases. 

Maria and Nikos, you talked about horizon scanning, not only managing what we think’s coming down the pipe, but maybe managing things we’re not seeing around the corner using this powerful tool of AI. Don, very good example of a use case of large volumes of seafood – 94% of all seafood that’s imported into the United States. So 94% of all seafood that’s consumed is imported. And how we can use this tool to ensure that seafood is safe for American consumers. Then, Cronan, you gave a use case of leafy greens, a recurring vehicle of foodborne disease. That’s really, really encouraging to me. 

AI Can Be a Team Sport

The third takeaway I heard was that, Nikos, I think you persuaded me, AI can be a team sport. I’ll be candid with you. When I got into this, I was thinking, well we’re going to hear about how AI can be used in my company or my specific supply chain. And what we’ve heard is maybe we should pause and think about it. 

We often say food safety is a collaborative effort. It’s a shared food system. But can we democratize big data in a way that the public and private sector share information or private to private sector, you have more data together and we all went on this and safety issue together. We’ve said food safety is not a competitive issue. So number three, I can be a team sport and I would challenge listeners to think about it a little bit unconventional like this and out of the box.

Focus on More than Just Technology 

My fourth takeaway was that you need to focus on more than just technology. And Maria, I appreciate you emphasizing the human element, which is the reality is AI isn’t going to replace the subject matter experts that are listening here today. We’re always going to need the best and brightest food safety professionals, but AI can be a powerful adjunct.

And then also, if you’re going to leverage it in your places of employment, it’s going to probably mean you change the way you work, so engage others in your organization and not just the IT department. 

AI in Food Safety is Already Here 

And then lastly, I heard, five, it’s already happening now. It’s already happening now. So let’s get started. I think we heard some powerful use cases on how it’s already being used in food. We all know if you have a smart device, the power of AI is in your smart device. If you’re online shopping platforms, AI is used there. If you’ve gone to a medical healthcare provider, power of AI in healthcare, and it’s happening in food too. And so let’s get started now using AI to strengthen food safety. 

Listen, I thought today’s session was fabulous. I don’t know about you. I want to thank all of our listeners for giving us your time. I hope we answered your questions, but rest assured there’s going to be more conversations about AI and food safety in our discipline and in our future.

I also hope this podcast gave you some real practical ways and ideas on how you can add AI as a powerful tool to your food safety toolbox. 

If you’ve enjoyed today’s podcast, I ask you to please take a moment and visit our Tech Talk podcast page on FDA.gov for updates on the next episode. 

Thank you all for listening.

Until next time, stay safe.