5

Potential Applications of AI to Large-Scale Food and Nutrition Initiatives

ADVANCING AI AND ML IN BIOMEDICAL RESEARCH AND HEALTH CARE AT NIH

Chris Hartshorn, chief of the digital and mobile technologies section in the National Institutes of Health (NIH) Clinical and Translational Science Awards program, discussed some challenges regarding artificial intelligence (AI) and machine learning (ML) for biomedical applications. He noted that biomedical research using ML or deep learning (DL) has accelerated rapidly over the last decade based on the literature and the number of investigator-initiated research proposals NIH has received. Another important trend is the growth, particularly over the last 7 years, in regulatory filings and U.S. Food and Drug Administration (FDA) approvals as a sign of AI and ML methods transitioning to the clinic. As of July 2023, FDA has approved nearly 700 AI/ML tools, with radiology applications accounting for 75 percent of the approvals. Hartshorn said that AI/ML tools have leapfrogged other new technologies in their path to clinical use.

Hartshorn raised the question of why dedicated public funds are still needed to support AI/ML application development, considering the accelerating translation to the clinic. One answer, he said, is that only a small percentage of the approvals used prospective data to support their request for approval (Wu et al., 2021a). Out of 130 approved AI/ML applications that one investigator examined, only 37 leveraged data from more than one site, and only four used prospective data. None of the four were for devices considered high risk. These findings, said Hartshorn, highlight some of the obvious problems with regulating these tools and the work needed to drive them to clinical utility and having measurable positive effects on clinical outcomes. Furthermore, this problem can be addressed via new approaches to clinical trials and the data produced being more AI ready or intended for an AI system and designing the trial accordingly.

Considering what is known about challenges in leveraging AI/ML for biomedical data, the data supporting approvals can be biased or incom-

plete, reducing the utility and accuracy of the clinical decision support (CDS) from AI/ML algorithms, which can have severe consequences if clinicians use them incorrectly, said Hartshorn. This raises other concerns about the little to no data on clinicians’ usage; for example, are they ever being used for clinical decision-making versus the intended use as support tools?

Hartshorn posited several other aspects as to why this field would continue to necessitate a large body of research. The uses of AI are not focused on health across the life span as a continuous and stochastic process, and most research is narrowly focused on optimizing flow and increasing accuracy of clinical decisions for single data types, such as most approved CDS tools for radiology. Thus, the impact has largely been confined to hypothesis confirming rather than hypothesis generating, which, from his perspective, is where AI/ML will ultimately be transformative for medicine and health care writ large. “Where I think where it will be truly transformative will be its use in multimodal, disparate, big data analysis, and dot-connecting, as these are tasks, simply put, [that] humans cannot do nearly as well or at all,” said Hartshorn.

The transformational aspect of AI and ML tools, he said, will be that they shift the biomedical research calculus from simply analyzing a data type and hypothesis testing to generating hypotheses from the evidence. “The ability to capture insight from the totality of medical evidence available at the individual to population level will entirely change how we deliver health care and ask research questions, yet this is by far the most challenging task, presenting a host of additional challenges to it ever being realized,” said Hartshorn.

One challenge is that biomedical data are multiscale, with both spatial and temporal dimensionality, Hartshorn noted. Linking these scales is not trivial, considering that for every data scale model and link, it would be necessary to understand, mechanistically, why an AI/ML algorithm is providing a particular answer for it to have any value. Moreover, each scale has unique temporal qualities, including its inherently stochastic biological nature and also an individual’s own life decisions, the decisions of others, and the environment in which they live—these present challenges to generating plausible correlations using AI/ML. As a result, all the ML and AI methods can point to correlations, but other multiscale modeling and in vivo experiments must be used to identify causality of correlations made by any AI system.

Hartshorn said that the second challenge is that within each scale, the data are multimodal and the tools for any individual measurement are typically not the same or even from the same vendor. In theory, AI systems have been used to generate correlations and predictions for integrated multimodal datasets before, but mostly in complex nonbiological systems. However, much remains to be done to help accelerate this for complex biological systems, said Hartshorn.

Further challenging the utility of AI in biomedical and clinical research using big data is that applying AI to biological questions requires models that can handle longitudinal, noisy, and incomplete data and a host of other aspects that are important to consider. For example, the provenance and standardization of measurements used to acquire patient data; privacy issues; categorical, nominal, and ordinal data; qualitative discrete and continuous data; and structured, semistructured, and unstructured data. Until recently, most biomedical data have not been generated with AI in mind, so little prospective data can be leveraged retrospectively. In addition, the sample size for training AI systems is a long-recognized pitfall, given that clinical trials typically do not recruit to the requisite sample size or demographic depth. How to mitigate sample size bias, such as “synthetic” data generation/injection, is still a nascent area.

Hartshorn said that with all the challenges that remain to push AI/ML to be a transformational tool in medicine, NIH is actively creating programs and funding to incentivize developing AI/ML approaches with more realistic, real-world utility in order to drive the evidence base and provide the answers to some of the challenges he noted. For at least 7 years, NIH has had many initiatives, large and small, focusing on AI/ML for biomedical research and medicine. These include the Common Fund’s Nutrition for Precision Health (NPH) program, Bridge to Artificial Intelligence (Bridge2AI) program, Office of Data Science Strategy’s (ODSS’s) Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD) program, and National Cancer Institute (NCI) and Department of Energy collaborations on multiple AI/ML-based programs. Bridge2AI, for example, aims to set the stage for widespread adoption of AI and tackle complex biomedical challenges by generating flagship datasets, preparing a road map for “AI/ML-friendly” data, emphasizing ethical best practices in the use of AI/ML, and promoting forming and training diverse teams. This program is developing automated tools, standardizing data elements, creating cross-training materials for workforce development, and disseminating products and best practices. Data generation projects cover topical areas, such as precision public health, functional genomics, salutogenesis, and critical care informatics. Furthermore, a myriad of smaller AI/ML programs have been established along with an ever-expanding number of successful investigator-initiated applications to NIH institutes and centers.

After briefly describing the NPH program, he mentioned that AIM-AHEAD¹ will establish mutually beneficial and coordinated partnerships to increase the participation and representation of researchers and communities underrepresented in developing AI/ML models and enhance

___________________

¹ Available at https://datascience.nih.gov/artificial-intelligence/aim-ahead (accessed January 9, 2024).

the capabilities of this emerging technology, beginning with electronic health record (EHR) data. Hartshorn encouraged people to go to the NIH ODSS website² for a continual feed of NIH’s active AI-specific and related funding opportunities. He also highlighted NIH’s decade-long collaboration with the National Science Foundation and Smart and Connected Health Program, which funds health-related aspects of AI and other synergistic areas in digital health.

THE NIH NPH PROGRAM

Holly Nicastro said that the goal of NPH is to develop algorithms to predict individual responses to foods and dietary patterns using data collected from 10,000 participants from diverse backgrounds on their physiology, metabolome, microbiome, genome, dietary intake, demographics, health history, psychosocial factors, behaviors, and environment. The idea is to use AI to identify the factors that explain why individuals or subgroups respond the way they do to certain foods or ways of eating. The study has an observational component that will generate data on the foods people eat in their everyday lives and an interventional component that will randomly assign participants to three dietary patterns, in their normal environments or a domiciled setting. Nicastro noted that NPH has participant ambassadors that inform the things that they would like to see studied and what they want to learn about their health.

As part of the All of Us Research Program, NPH will have access to its data and be able to study the techniques that Edward Sazonov, Rob Knight, and Susan McRitchie described and merge various data sources. For example, aligning the timing of food intake with metabolite changes may provide information on how the metabolome changes over time.

Privacy and trust are concerns, said Nicastro, given that some studies will ask participants to wear cameras that may capture other people. “We also need to be aware of the increasing possibility of identifiability with the data we are collecting,” she said. “We are building more and more toward having digital twins or digital avatars of ourselves, so we are starting to blur the boundaries of what could be considered personal identifiable information.” In addition, as this work may provide information on disease risk, it will be important to remember that this does not equal disease and exercise caution about any potential stigma or insurance-related issues.

Nicastro said that the program does not want to identify and troubleshoot issues of privacy and trust in real time. “We want to be taking proactive approaches to build trust versus assuming that we are starting with that trust and then losing it,” she said. Regarding diversity, equity,

___________________

² Available at https://datascience.nih.gov/artificial-intelligence (accessed January 9, 2024).

and inclusion (DEI) and accessibility, she sees a tremendous opportunity for advances in data collection to engage people who have not participated in nutrition studies before and ensure that any findings are relevant to and reach diverse populations. “We all need to be considering how any of these big data findings can be translated and implemented in clinic and community settings and not just for those with disposable income,” she said. One possibility is developing simple point-of-care technologies, perhaps a finger prick, that can capture a simple metabolomic signature to better inform personalized nutrition recommendations.

Nicastro is also excited about the ability of passive image-assisted technologies to provide a fuller picture of what participants are eating to complement self-reported behavior. As technologies such as smart toilets or smarter fecal sample collection devices improve, they may address barriers, such as people’s reluctance to collect fecal samples or store their biospecimens in the refrigerator before their clinic visit.

Collaborations will be critical to moving the field forward, said Nicastro. “We need nutrition scientists in the same room with the engineers, but beyond bringing people together, we need cross-training of the workforce,” she said. Regarding where the field might be in 5–10 years, she wondered if it is overpromising or being realistic. “Are we going to walk into a restaurant, pull out our phone, access our digital twin, and maybe simulate what will happen to our cardiovascular risk profile if we order the burger versus the salad? Probably not,” said Nicastro. But she is certain that initial results from NPH will power algorithms that provide a better understanding of why some people respond in certain ways to food.

The immediate next steps, said Nicastro, will be to validate algorithms in different populations and settings and conduct trials to see if targeted guidance based on these algorithms produce the desired results. That could lead to including predictors and the evidence behind them in the Dietary Guidelines for Americans, which would be a great example of precision nutrition. The idea, she added, is for practitioners to use this information to inform dietary advice, whether that is the calorie level an individual might need or the foods they should eat.

Tempering her excitement about the future of precision nutrition are the realities of health care delivery today, where physicians are spending less time with patients while being asked to explain ever more complicated topics, such as computer-generated risk profiles. “When we start to add in more of these personalization or precision factors for a patient, is this going to be more burdensome for the practitioner and the patient, or will knowing that this advice was tailored to that individual serve as a tool to make them better empowered to stick to their plan?” asked Nicastro. “We will find out.”

ETHICS, ACCESS, LEGAL FRAMEWORKS, AND FAIRNESS IN AI

Janie Hipp, inaugural president and chief executive officer of Native Agriculture Financial Services, began by noting that agricultural producers are facing unprecedented complexity in decision making and that agricultural AI could support, but not necessarily simplify, many of those decisions (Sørensen et al., 2010). Agricultural production, she added, is a data- and algorithm-intensive process, and AI encounters every challenge that speakers have discussed about dealing with big data systems (Ayoub Shaikh et al., 2022). Analyzing and absorbing the copious amounts of data into farm production systems will be essential to achieve a fair, just, and efficient system of food production in the United States. “We cannot get this wrong, or we are headed for a picture of a very unfair, unjust food system writ large,” she said.

Hipp explained that approximately 6 percent of agricultural producers account for 60–70 percent of the nation’s agricultural output. The other 94 percent fall in the small to mid-size category³ but have the same challenges as large-scale producers. However, the needs of the farm manager and ranch manager will be different based on the size and nature of their operations.

The architecture of smart agriculture is complex, and farmer decisions are driven by environmental, market, and cultural conditions and any available data (see Figure 5-1). The farm household, said Hipp, is where human health and decisions around food, food preparation, consumption, and nutrition collide with what is happening on the farm as a component of the farm system. “That is a very interesting place to be, because how you believe farmers actually respond to data can be quite different in real life than what you think it is going to be as you are researching this arena and trying to contemplate their reactions,” said Hipp.

Hipp listed several questions regarding the role of government regulation and safeguards:

• Who owns the data that will drive these innovations?
• Who owns the innovations derived from these data?
• Who is responsible for ensuring the veracity of the AI systems?
• Who is responsible for security of the AI systems?
• What is the legal framework that surrounds AI?
• What is the ethical framework that surrounds AI?

As Hipp noted, agriculture is inherently data rich, with farmers generating data every single day from the land, their equipment, their inputs,

___________________

³ Available at https://www.ers.usda.gov/data-products/chart-gallery/gallery/chart-detail/?chartId=108013 (accessed December 29, 2023).

and their labor. How they do so will depend on whether they own the land and operation or just manage or work on it. Many people involved in farm production have questions about the producer’s ownership rights associated with an AI model if it is built with data the producer owns. For example, who owns it, at which stage in the creation process does ownership or shared ownership attach, and where does the government come in during this process? What happens when the data used to create the AI model are bad? How do bad data affect the farmer whose raw data go into the AI?

Another important question concerns the national security implications of AI in agriculture, given that food security is national security and AI will affect food security, food access and availability, and success of the sector. For example, what aspects of data and AI in agriculture have national security implications? Those implications lead Hipp to believe that government must be involved. “We have to get this piece of the puzzle right,” she said.

One problem Hipp has seen as a lawyer is that farmers and ranchers sign nonnegotiable contracts to obtain the equipment and supplies they need to continue their operations. What this means is that although 77 and 80 percent of them are worried about data security and believe they own their own data (Yu et al., 2021), respectively, they actually do not according to the contracts she has seen and analyzed. “Ownership and control of farming data is a significant concern for me, and it should be at the policy level,” said Hipp. “It allows for market speculation and control of ag operations, and it can start to spin off into national security issues very quickly.” It also feeds into her concern about bad data.

As an example of the harm that a farmer can experience from bad actors using their data, Hipp explained that farmers regularly participate in the agricultural census and with research communities in on-farm demonstration projects that generate data for use by other farmers and ranchers. In the 1990s and 2000s, data mining resulted in environmental lawsuits directed at farmers who had for decades poured their data into publicly accessible datasets and did things in a certain way based on feedback from the research community. Despite identity protection laws, the small set of farmers makes it fairly easy to figure out who generated the data.

Hipp explained that the legal mechanisms for dealing with how U.S. law protects data in the realm of AI do not exist. In her opinion, patent, trademark, and copyright laws are not set up to deal with the reality of AI. She explained that U.S. law classifies data as facts (Yu et al., 2021). As the basic facts underlying certain agricultural commodities, farm data lack a creative element that can be defined as an intellectual property whose ownership could be protected by copyright laws. Therefore, legally speaking, farmers may not own their raw data. Hipp noted that patent, trade-

mark, and copyright laws are controlled at the federal level, and trade secret issues are primarily controlled by state laws.

States have passed laws to protect data at the farm gate level so they could not be used by outside forces against the farmer or rancher in a way that affects what they do. What concerns Hipp is a possible patchwork of state laws that will cause an uneven growth trajectory for AI.

Hipp said that the driving force for her work and involvement has been that agricultural production improvement has always been about stemming the tide of hunger and malnutrition, although agriculture is not just an economic act but an act of community, sacredness, and protection of one another. However, it can be a means of exploiting communities and people. That raises the issue of the interaction with and impact on the AI model of the most underserved actors. Hipp said that it is hard to consider ethics in AI without confronting the nebulous concepts of community, sacredness, and protection. “I fear for us if we do not insist that these unseen forces for why people make the decisions that they do are a part of the conversation,” she said.

The other piece of this puzzle that troubles Hipp is that data and AI/ML applications are only usable if one has the time and space, and the larger players are more able to actually absorb data and applications into their operations and make decisions with that data. “But if you are at a smaller end of an agricultural operation, you are many times operating on knowing your gut and what is happening right at that moment. You do not have the luxury or the time or the money, literally, to ponder the data,” said Hipp. This issue raises questions:

• How to target AI both at the larger players and simultaneously at the vast numbers of smaller players?
• How to stabilize access to data and information?
• How to ensure AI is size specific?
• Is the goal to improve the lives of the entire community or only a limited set of people?

Hipp added that if it is too hard to consider ethics in AI, the nation must be prepared to live with the outcomes.

Legal frameworks, said Hipp, must be reformed at the same time society embraces its AI future in agriculture. Questions that need answering include the following:

• What is the regulatory regime, and who holds it?
• How to ensure privacy of data generated on the farm?
• How to ensure permissions/authorizations for using data occur, and if compensation for data is provided, how do we ensure the amount is fair?

• How to focus on incentives in AI and not just regulation? Are incentives the only approach that allows for the continued presence of small to mid-sized players?
• What are the unintended consequences to various types/sizes of producers as AI proceeds?
• How to ensure informed consent for using data?
• A solely “regulatory” approach to the output of AI will further ensure uneven distribution of AI. How do we safeguard against this?
• AI can affect environmental management, risk management, economic analysis, resiliency, and adaptation to climate change; which of these have the greatest potential for harm and need the greater level of regulation and scrutiny?

Hipp said that in her opinion, the U.S. Department of Agriculture (USDA) is the only informed choice regarding regulation of AI and agriculture. However, USDA does not oversee regulatory regimes and deals more with incentives rather than regulations.

Not considering ethics, access, legal frameworks, and fairness in AI will lead to exploitation, said Hipp, and prevent accessing inherent knowing and sacredness, which is a huge issue for Indigenous communities such as hers. “There is a deep, deep understanding within Indigenous communities and tribal governments in the United States that data sovereignty is important, and we are very highly concerned about these issues,” she said. The conversation about data ownership and AI is just starting in the native producer community.

Other pitfalls and challenges associated with a failure to consider ethics, access, legal frameworks, and fairness in AI include the difficulty of recognizing and quantifying the harm caused; the rise of state-level pushback that will lead to a patchwork of policy and an uneven policy landscape; determining who has the right to clean up the data and how to deal with bad data and bad actors; and the question of who is at the table throughout the discussions to embed fairness in AI policies and who is creating the decision support systems that allow producers to stop drowning in their data and start using them. Datasets in agriculture production are almost unusable by the farmers who generate the data and can experience the most improvement in the resiliency and viability of their operations, said Hipp.

Hipp pointed to the need to attend to the DEI, accessibility, and fairness of AI systems as they are being built, modified, and enlarged. Competent lawyers are needed at the table now, as is impressing upon lawmakers at every level that this train is moving fast and, without a focus on achieving the proper balance today, agriculture is headed for a rapid exacerbation of

exclusion, exploitation, lack of privacy, and failures of control. Finally, she said, “If we have any hope of a fair and just food system that takes into account national and food security and that achieves the highest good for the most people, we must get this right.”

MODERATED DISCUSSION

When asked if he sees any pitfalls regarding the program NIH is funding on the use of AI in nutrition (NPH program), Hartshorn said that he does not, given this is discovery-based versus strictly hypothesis-driven science—unique to traditional NIH efforts. Hipp largely agreed with Hartshorn but also wondered if advances in biomedicine and nutrition will send market signals back into the food system writ large and change what farmers grow, how and where they grow it, and even who grows it. The challenge, she said, is making any transitions as fluid and least harmful as possible and providing incentives that allow farmers and ranchers to be full participants in the AI arena in a way that does not exacerbate concentration in agriculture.

Hipp said that her vision for the future is to have a tool that a farmer standing in a field and wondering how to deal with a certain circumstance can use to get various options that will guide them in a way that does not harm their operation. In addition, all producers, regardless of their size, will have access to the same set of decision support tools. That would allow small and mid-size producers to have a viable path forward and stay on the land and continue producing food. She added that as an Indigenous person, it is important that whatever gets developed, it must acknowledge that she is different, that each person is an individual.