Proceedings of a Workshop—in Brief

The Role of Advanced Computation, Predictive Technologies, and Big Data Analytics in Food and Nutrition Research

Proceedings of a Workshop—in Brief

The Food and Nutrition Board of the Health and Medicine Division of the National Academies of Sciences, Engineering, and Medicine (the National Academies) held a workshop, The Role of Advanced Computation, Predictive Technologies, and Big Data Analytics in Food and Nutrition Research, on October 10–11, 2023, in Washington, DC. The workshop explored opportunities and challenges related to applying advanced computation, big data analytics, and high-performance computing, including artificial intelligence (AI), machine learning (ML), and deep learning (DL), to support advances in food systems and nutrition research. Speakers discussed definitions and methods; the appropriate use of evidence generated from these methods to inform food- and nutrition-related programs and policies; considered issues related to diversity, equity, inclusion, bias, and privacy; identified opportunities and challenges related to capacity building and training; and explored the future potential of these methods in food and nutrition research. The workshop sessions highlighted applications and lessons learned from studies of AI, ML, and DL methods in both food and nutrition research and other fields.

This Proceedings of a Workshop—in Brief highlights the workshop presentations and discussions and is not intended to provide a comprehensive summary of information shared.¹ The information summarized here reflects the knowledge and opinions of individual workshop participants and should not be seen as a consensus of the workshop participants, the Food and Nutrition Board, or the National Academies.

SETTING THE STAGE

Introductory Remarks

The workshop began with introductory remarks from three speakers and three presentations laying the groundwork. Patrick Stover, director of the Institute for Advancing Health through Agriculture at Texas A&M University, noted that this is a critical time for nutrition science, agriculture, and public health, with dietary patterns being a major driver of rising health care costs. Including health and disease reduction as goals for food and agriculture will require transforming advances across the entire food and agriculture value chain, he said, and advances in AI could help identify the causal relationships that underpin connections among agriculture, food, and health.

__________________

Cindy Davis, national program leader for the U.S. Department of Agriculture (USDA) Agricultural Research Service Human Nutrition Program, explained that its mission is to define the role of food and its components in optimizing health throughout the life cycle for all people by conducting high-national-priority research. Understanding the complex interactions within the food system related to human health requires multidisciplinary teams that assess inputs and effects from all sectors. AI and ML, she said, have shown promise in many applications to better understand and predict interactions between food and nutrition-related data and health outcomes, particularly when large amounts of data need to be structured and integrated.

The increase in diet-related chronic diseases is complex with multiple etiologies, said Davis. She noted that the field’s understanding of the food-related physiologic processes underlying health and disease prevention is expanding constantly but that further advances will require new information relating to how dietary patterns, specific foods, nutrients, bioactive components, and physical activity influence these processes. In addition, emerging evidence suggests that many subpopulations have differential responses to diet and chronic disease risk and that the large, interindividual variability and individual responses to diets and environment are not well characterized.

Jennifer Tiller, deputy staff director for the House Committee on Agriculture, said that the committee chair has stated that he believes that policy should use the best science—not political science—and has called for improved nutrition policies that can mitigate increasing instances of diet-related chronic disease among the populations served by the programs the committee authorizes, including the Supplemental Nutrition Assistance Program, which serves over 41 million people. Previous testimony before the House committee explained that the right resources, research, data, modernized programming, technology, and appropriate and effective federal dietary policy will enable USDA, the states, communities, and academia to improve the nutrition of the millions of people who rely on this program.

AI and Health

Anant Madabhushi, Robert W. Woodruff professor and research neuroscientist in the Departments of Biomedical Engineering, Radiology, and Imaging Sciences and Biomedical Informatics and Pathology at Georgia Tech and Emory University, said that AI has implications for a wide range of areas, including health care, education, and agriculture. However, he emphasized that AI is not magic, and it needs to be used thoughtfully and intentionally when developing algorithms. Interpretability, affordability, and equity are key considerations.

Madabhushi listed key considerations for developing AI tools for precision medicine and nutrition. The first is interpretability and developing handcrafted, engineered approaches that start with a set of attributes or features with more inherent interpretability attached to them. He urged caution around using black box models for decision making due to their lack of interpretability. A second consideration is affordability and taking advantage of routinely acquired data such as radiologic scans and pathology images, particularly in low- and middle-income countries that may lack the ability to acquire sophisticated data or sophisticated technology to acquire more data. The third consideration is equity and whether an algorithm works across the plurality of populations. He also noted the importance of being intentional in the development of more representative algorithms.

As an example of the need for interpretability in health care, Madabhushi described a data-driven DL approach that he and his collaborators used to create an algorithm that can detect malignant cells in breast pathology images. They used the same approach, in a study he discussed, to create an algorithm from images of annotated endomyocardial biopsies that can detect heart failure from unannotated biopsies. The algorithm’s predictions were 97 percent accurate, and expert pathologists were only 74 percent accurate. However, when the algorithm analyzed another set of biopsies, the accuracy dropped to 75 percent. The reason was that the biopsy slide scanner received a software upgrade that changed the slides’ appearance so subtly that humans could not tell the difference, but it was significant enough to affect the algorithm’s performance. Both of these are examples of uninterpretable, or “black-box,”

models created directly from data by algorithms that make predictions without a clear picture of how they combine the data to make those predictions.

Promises and Challenges

Judy Gichoya, associate professor of interventional radiology and informatics at Emory University, said that applying AI to predictive nutrition will be a bigger, more difficult task than using AI with imaging data. She noted that although ML algorithms can provide a great deal of information about patients, they can perpetuate bias and disparities when deployed widely if the models are trained using data from a homogenous, unrepresentative population. For example, she discussed a study on four DL models designed to segment cardiac magnetic resonance images that were all found to have performance biases depending on whether the images came from a man or woman or a Black versus a White individual. In another instance, she reported on a study of a DL model for analyzing chest radiographs that severely underdiagnosed underserved patient populations, including women, Black individuals, and people on Medicaid.

Gichoya noted serious issues regarding bias and AI systems because investigators can introduce bias in the data depending on how they assemble the dataset they use to train their models. For example, in imaging studies the investigators may not image all eligible patients, only use inpatient images for training an AI that will be used on outpatients, or inadvertently include mislabeled images that skew the resulting dataset.

There are also serious questions regarding privacy and regulation of AI systems, posited Gichoya, such as how and when is consent needed for data sharing, are institutional review boards empowered to protect patient privacy in the era of AI, is it possible to sufficiently deidentify and anonymize patients, can patients opt out of sharing their data, and is there a mechanism to obtain partial consent, for example consent for lab results in an electronic health record versus no consent for sensitive patient images.

Benoît Lamarche, professor in the School of Nutrition at Université Laval and scientific director and founder of the Nutrition, Health and Society Center, said that applying AI/ML in nutrition research has great potential to increase the capacity to manage and analyze big datasets and perhaps identify new relationships and patterns in the diet. However, he pointed out, this is difficult because of the complexities and challenges in measuring relevant characteristics and relating them to the behaviors and societal factors that influence how and what people eat.

Historically, nutrition research has established correlations rather than causation, but the potential for AI is to develop some ability to predict outcomes that can enable preventive interventions. In particular, he said, AI can advance nutritional science by taking individual data, connecting these data with the variables that influence individual choice, providing a better understanding of dietary patterns, enabling the prediction of health outcomes, and identifying actionable variables or features that public health officials can use to educate the public and change behavior. He noted the importance of bringing together researchers from the many disciplines (e.g., engineers, modelers, nutritionists), that will be involved in using AI/ML in food and nutrition research and developing a common culture. Such a common culture, he stated, would include a shared vocabulary, standardization of methods, capacity building, and training.

APPLICATIONS AND LESSONS LEARNED

Edward Sazonov, the James R. Cudworth endowed professor in the Department of Electrical and Computer Engineering and head of the Computer Laboratory of Ambient and Wearable Systems at the University of Alabama, said that the first and simplest task to perform with a wearable device is detecting when someone is eating and perhaps how much they are eating. One approach is to develop wearable devices that record the timing and duration of eating, evaluate daily eating patterns, and evaluate the microstructure of eating.

For infants, monitoring sucking or swallowing would be effective means of detection, and hand gestures or chewing in a slightly older child would more applicable. For older children and adults, hand-to-mouth monitors could provide information about eating but would have relatively low accuracy because of the similarity to

everyday gestures and offer no insights into the type of food consumed, said Sazonov. Swallowing monitors worn around an individual’s neck would detect food and beverage consumption, but people do not like wearing such devices, and they also do not identify the food type. Chewing monitors are reliable for detecting food and beverage consumption, although sipping beverages may generate false negatives and chewing gum or lip biting may generate false positives.

Some wearables, said Sazonov, can take pictures of what the wearer is eating. A device his group developed attaches to an eyeglass temple and captures images of the wearer’s food, reveals the eating progression, and automatically detects eating by capturing images of food intake. When an ML algorithm running on the device decides that a person is eating, it takes a picture of the food. A nutritionist or image recognition system analyzes the images to identify the type of food and portion size, although the error rate for both can be large.

Wearables are good for passive detection of eating events, Sazonov summarized, but do not provide complete insights into the consumed food, even when combined with imaging, as these are still indirect measurements. Additionally, he noted, ethical and privacy issues need to be addressed.

An individual’s microbiome is intimately linked to diet. Rob Knight, the Wolfe Family endowed chair of microbiome research at Rady’s Children’s Hospital of San Diego and director of the Center for Microbiome Innovation and professor of pediatrics, bioengineering, data science, and computer science and engineering at the University of California—San Diego, said that the dramatic reduction in the cost of DNA sequencing has enabled many new applications to the microbiome and generated an enormous amount of data. With these data comes the ability to build predictive models for many traits related to health from the microbiome.

Knight said that ML has been used to relate the microbiome to nutrition. He discussed a study that put 800 people on continuous glucose monitors and measured a variety of factors that led to individualized glycemic responses when given a defined sequence of meals. This enabled the researchers to isolate the effect of different foods on blood glucose response after eating and show that the individualized glycemic response for a particular person to a particular food varied greatly. One surprising finding was that for a sizable fraction of the cohort, rice was worse for their blood glucose than ice cream, and the microbiome was the best predictor of who was in which group. These findings, Knight stated, suggest that, when it comes to individual glycemic response, it might be possible to change someone’s microbiome to put them into the category of people who should eat ice cream instead of rice. The idea of changing the microbiome is being explored in other health outcomes, such as cancer.

Susan McRitchie, lead biostatistician and program manager in the Metabolomics and Exposome Laboratory at the University of North Carolina—Chapel Hill Nutrition Research Institute, explained that an individual’s metabolome represents the biochemical fingerprint of low-molecular-weight metabolites present in tissues or biological fluids and is ideal for studying precision nutrition and precision medicine. The metabolomic signature comprises endogenous metabolites reflecting host and microbial metabolism and exogenous metabolites derived from external exposures throughout an individual’s life. Metabolomic studies, said McRitchie, can be hypothesis or data driven and use ML to identify patterns in the data. Metabolomics data can be analyzed using a wide variety of ML methods, including random forests, neural networks, and even some DL models. The data can provide mechanistic insights from pathway analysis that can identify disease biomarkers, pharmacologic targets, nutritional interventions, and genetic links to disease. Exposome research, she added, is key to informing precision nutrition by providing an understanding about how exposures and perturbations in a person’s metabolism are linked to their genetics and states of health and wellness. Pathway perturbations, she said, can help identify potential targets for nutritional and pharmacologic interventions.

Responders and nonresponders are important in precision nutrition, McRitchie noted, because individuals

have different nutrient requirements and can have different responses to nutrient intakes. An individual’s response is going to be related to their exposures, health status, and their genetics, she stated, and studying how people respond allows for discovery of mechanisms and identification of nutritional targets.

One study McRitchie discussed used metabolomic and microbiome data to identify biomarkers for and pathways involved in osteoarthritis. Untargeted analysis of fecal samples from individuals with obesity and with or without osteoarthritis identified several factors that could predict osteoarthritis: metabolic disruptions related to intestinal permeability, perturbations related to precursors of polyunsaturated fatty acids such as omega-3, levels of the short-chain fatty acid propionate, which is produced by microbes in the colon and associated with dietary fiber intake, and levels of glucosamine. Using a data-driven analysis, McRitchie and her collaborators also identified a link between osteoarthritis, gut microbes, and environmentally relevant metabolites, inflammation markers, and endogenous and microbial metabolism.

NUTRITION RESEARCH

AI has great potential for the nutrition field, Lamarche stated, including increased capacity to analyze and manage large data sets, dietary assessment, prediction of outcomes, and social media content analysis. However, he further stated that there are significant challenges to address, such as appropriately considering data quality, developing a common language between nutrition and computational science, and establishing standardized methods and approaches. As an example of an AI-based application, Lamarche discussed a tightly controlled feeding study in which men and women with metabolic syndrome ate a North American diet for 5 weeks and then a Mediterranean diet for 4 weeks to determine the effect on their metabolome. The goal was to develop an ML algorithm that classifies people according to diet based on their untargeted metabolomics data.

The resulting algorithm was 99 percent accurate, far better than the typical published study, and it showed the value of having full control of the diet, which generates a much cleaner dataset. However, when repeating the experiment with self-reported dietary intake data, Lamarche found more noise in the dataset, and the accuracy of the same algorithm dropped by almost 20 percent. The lesson, said Lamarche, is that the signature may not be accurate when using either untargeted or targeted metabolomics without carefully controlling the diet. This is an important issue because most published nutrition studies rely on self-reported dietary intake data.

For precision nutrition, the idea is to use all the data available from one individual to predict a health outcome for them, said Lamarche. One such study he discussed put prediabetic individuals on either the Mediterranean diet or a personalized diet based on information from microbiome data, blood tests, questionnaires, anthropometrics, and food diaries. Creating the personalized diet relied on an ML algorithm that integrates these data to predict glucose responses after a meal; the mean daily glucose levels and after-meal glucose levels were significantly lower in that group, showing that using more data to personalize the intervention was working. The challenge, said Lamarche, is figuring out how to implement this type of intervention in clinical practice.

As an example of how AI can benefit precision public health, Lamarche discussed a study that characterized and mapped local environments by the relative proportion of low-quality food offerings and compared that to a map of socioeconomic status and dietary habits from Web-based dietary intake data. The ML model developed from these data could identify where people have a low-quality diet and live in a low-quality food environment. From a public health perspective, this type of mapping model can identify those areas where the environment is unfavorable to high diet quality and develop interventions that change the environment of those areas. One caution, said Lamarche, is that these maps could lead to stigmatization of different neighborhoods.

Sai Krupa Das, senior scientist on the energy metabolism team at the Jean Mayer USDA Human Nutrition Research

Center on Aging and professor at the Friedman School of Nutrition Science and Policy at Tufts University, said that the growing intersection of AI and nutrition research provides exciting opportunities for streamlining conventional research protocols and revolutionizing clinical nutrition applications. In Das’s view, much of the difficulty in crafting personalized nutrition recommendations lies in understanding the interplay among internal (e.g., microbiome, genetics) and external factors (e.g., diet physical activity) that produce variations across individuals. Current approaches to address these interactions are primarily in fields such as genomics, proteomics, and metabolomics, but a comprehensive understanding of diet and disease relationships requires the type of integrated, systems-wide approach for which AI/ML methods are well suited. Das noted that over the past 2–3 decades, data have been collected on a variety of factors, but the challenge has been that such data exist in silos.

An absolute must, said Das, is taking a team science approach for generating AI-ready data. Adapting to the widespread use of data-driven technologies will require supporting a team science approach and developing professionals with clinical and computational skills. Engaging the research team in all aspects of a project, from inception and planning to completion, analysis, and interpretation of findings, is important, she said, and will help integrate experts in all relevant fields to provide insights from diverse perspectives. Das recommended engaging data scientists as integral peer collaborators. Multidisciplinary training and education are also critical for building interdisciplinary teams, as is emphasizing that AI is a tool to inform study designs and facilitate decision making, not a replacement for human experts.

SUPPLY CHAINS AND NUTRITION

Christopher Mejía-Argueta, research scientist at the MIT Center for Transportation and Logistics and founder and director of the MIT Food and Retail Operations Lab, and his collaborators are tackling the challenge of combating food malnutrition by creating innovative strategies to address the related problems of food waste or loss across the supply chain, from the field to the grocery store or farmers’ market. Malnutrition affects other things besides health, including economic growth, social development, productivity at work, and a country’s economic performance.

Mejía-Argueta and his team are also focusing on food safety and stakeholder behavior. He noted that technology works as an enabler, and it is important to provide good governance and policies that will be helpful for all population segments. Mejía-Argueta said that the main challenge in the food supply chain is perishability. AI and ML algorithms are embedded in the supply chain in the decision-making processes that connect the (forecasted and actual) order and the inventory being managed to the consumer. AI and ML may help predict how a consumer will behave, how changes in the supply chain can affect nutrition, and how supply shocks may affect the yield and other parts of the upstream value chain.

Elenna Dugundji, a research scientist at the MIT Center for Transportation and Logistics and collaborator of Mejía-Argueta, described a project to minimize cost and maximize the use and preservation of resources for a global airline catering company that combined data analytics and ML algorithms with a system dynamics framework. The team modeled the system with a 91 percent accuracy, evaluated eight different food waste management alternatives regarding the cost and environmental impact, and was able to provide the sponsoring company with actionable, data-driven recommendations to aid in developing its first attempt to create a global food waste management strategy.

Becca Jablonski, co-director of the Food Systems Institute and associate professor at Colorado State University, said there is always the potential of omitted variable bias or not considering the importance of relationships when building models. She tries to deal with these by grounding conversations in the community. Regarding missing data, she called for focusing more on getting the best data possible for tipping points that lead to greater impact (i.e., places where small changes can lead to a big difference). An opportunity to leverage unsupervised learning also exists, especially with missing data. For her work, getting

information along the supply chain is the hardest piece, she said.

Jablonski noted that a lack of data can make it challenging to support historically disadvantaged and underserved farmers and ranchers without exacerbating challenges, suggesting the federal government needs to rethink its data collection efforts. It may be fruitful, she said, for the government to invest in web scraping to get supply chain and farm gate pricing data, particularly given the growth in online platforms for farmers. She added that including behavioral components in models is important for ensuring that decision making is integrated in models in a meaningful way. It is crucial to show the evidence used to power a model to its intended beneficiaries and generate ideas that can improve the end user’s situation.

One of her major concerns about leading big modeling efforts is that she has not found a way to bring the best disciplinary science and robust methods to interdisciplinary projects. It is important, she said, to ensure that models reflect actual production systems in a given region. Regarding training students, Jablonski said that the challenge is balancing the trade-offs between depth and breadth of knowledge. Training students to be “T” thinkers (the vertical line is depth, and the horizontal line is their ability to communicate across silos), will be critical, although she does not believe that pure interdisciplinary programs are the answer because students need depth in at least one area.

BUILDING A MULTIDISCIPLINARY AND INCLUSIVE WORKFORCE

As touched on by other speakers, Saurabh Mehta, the Janet and Gordon Lankton professor, director of the Program in International Nutrition, founding director of the Cornell Center for Precision Nutrition and Health, and codirector of the National Institutes of Health (NIH)–funded Cornell Center for Point-of-Care Diagnostics for Nutrition, Infection, and Cancer, said that advanced training in AI for precision medicine is one component of building capacity in the field. To promote that, NIH issued a request for applications (RFA) aimed at building a future workforce capable of making pivotal discoveries using an increasingly complex landscape of big data and an array of data tools to tackle complex biomedical challenges in nutrition science and diet-related chronic diseases. One RFA element called for assembling teams of interdisciplinary scientists across nutrition, biomedicine, behavioral science, and computational methods.

The resulting training program he developed with collaborators at Cornell, Cornell Tech, Weill Cornell Medicine, and the U.S. Military Academy at West Point has one director, six codirectors, and 23 faculty members spread across nutrition, computational biology, neurobiology, medicine, population health sciences, computer and information science, and engineering, with plans to add expertise in ethics and fairness in AI. The plan is to have four predoctoral trainees and one postdoctoral trainee. The predoctoral trainees will be divided between those going for the Ph.D. in nutrition or a related biomedical field who will minor in computer science and those going for a Ph.D. in computer science who will minor in nutrition. He anticipates that the postdoctoral fellow will already have training in computational fields and want to apply that skillset to nutrition. A key aspect of the program is that it will require all five trainees to have two mentors, one focused on nutrition and the other on AI, and complete a capstone course that provides them with real data to analyze. Mehta said that if the program is renewed after its initial 5 years, the goal would be to add a second postdoc with expertise in nutrition who wants to learn to apply computational techniques to nutrition research.

One challenge the program faces is the need to be flexible and evolve constantly as the field advances and requires new expertise. Another challenge is that computer science courses can present a steep learning curve for many nutrition students. Coursework will need to be supplemented with cocurricular activities to develop interdisciplinary scientists, Mehta said, and responsible conduct-of-research training will need to be reimagined and strengthened as the nascent program develops to account for ethics, fairness, and equity in AI.

Angela Odoms-Young, the Nancy Schlegel Meinig associate professor of maternal and child nutrition and director of the Food and Nutrition Education in Communities Program and New York State’s Expanded

Food and Nutrition Education Program at Cornell University, addressed two key questions regarding diversity, equity, inclusion (DEI) and belonging:

• What is the significance of promoting DEI, along with justice and belonging, in applying advanced computation, big data analytics, and high-performance computing in food systems and nutrition research?
• How do we create a just, diverse, inclusive, and equitable training environment to support robust and ethical application of advanced computation, big data analytics, and high-performance computing in food and nutrition research?

She noted that the potential exists for AI to be effective at closing the gaps for the many U.S. subpopulations experiencing higher rates of diet-related morbidity and mortality than the general population, but may also exacerbate disparities. Although much of the work in this area focuses on race, gender, and income, similar disparities in health outcomes exists for members of the LGBTQ+ community and individuals with disabilities. Despite strong efforts at the national, state, and local levels, disparities in food insecurity and related outcomes continue to persist, she said.

Having diverse research teams is important because people with different lived experiences will pose different questions, said Odoms-Young. She quoted Timnit Gebru, who said, “If we don’t have diversity in our set of researchers, we are not going to address problems that are faced by the majority of people in the world. When problems don’t affect us, we don’t think they’re that important, and we might not even know what these problems are because we’re not interacting with the people who are experiencing them.” She reported on extensive evidence that indicates diverse teams can generate more inclusive and relevant research questions, which is important when considering the social and structural drivers that link to diet-related conditions and behaviors.

Diverse research teams can alter the behavior of a group’s social majority in ways that lead to improved group thinking, said Odoms-Young, and be less likely to be influenced by unconscious biases and stereotyping, leading to more objective and unbiased findings. Diverse team members can bring unique knowledge, skill sets, and subject matter expertise, which can enrich the process and result in a more comprehensive understanding of the subject matter, and their research is more likely to reach and resonate with a broader audience based on its relevance to various communities and stakeholders. She noted, too, that Indigenous knowledge can offer a different way of considering AI and how people relate to one another.

Odoms-Young cited numerous barriers to DEI and belonging. One barrier is implicit bias. Researchers and team leaders may have unconscious biases that influence their decision making regarding team composition and lead to underrepresentation in research teams, even when qualified candidates are available. Another barrier is limited networking opportunities, which are often crucial for career advancement in research. Lack of inclusive hiring practices can perpetuate a lack of representation on research teams, with traditional hiring methods favoring candidates from majority groups. Tokenism is another barrier, she noted, and can lead to feelings of isolation and marginalization and limit the potential impact of diverse perspectives.

Odoms-Young said that when thinking about DEI and belonging, intersectionality is important because all people, even those within certain subgroups, may not be the same depending on their intersectional identities and exposure to intersectional oppression. DEI training that involves people bringing their lived experiences is paramount. It can help people realize they have the will but not the tools to be inclusive and create an environment of cultural safety. She emphasized that it is important to both bring in a diverse set of trainees to prompt constant thinking about DEI and have trainers that are empowered and inspiring.

One piece of advice Odoms-Young gives to her graduate students is to try to implement the world they want to see. This may take time, she said, because it involves a culture shift, and it will require someone to emphasize

the need to build diverse research teams. This shift will require people in power to be intentional in making room for others; collecting data, tracking, and commitment to change; and cross-cultural mentoring to address the needs of students from diverse backgrounds who are not represented on the faculty.

Carmen Tekwe, associate professor of biostatistics at Indiana University at Bloomington, said that when her students develop a statistical method, she has them apply the method to different subgroups to see if the method performs equally well. So far, the methods have not, but exposing her students to the idea that statistical methods do not always apply to all populations introduces them to the concept of equity in research. She then has her students try to develop group-specific methods.

APPLICATION TO LARGE-SCALE FOOD AND NUTRITION INITIATIVES

Biomedical Research

Chris Hartshorn, chief of the Digital and Mobile Technologies Section in the Clinical and Translational Science Awards program at NIH, said that the transformational aspect of AI/ML tools will be that they change the biomedical research calculus from simply analyzing a data type and subsequent hypothesis testing, to generating hypotheses from all the available medical evidence, and that will change how we deliver health care and ask research questions. He noted that based on the research literature and the number of investigator-initiated research proposals NIH has received, biomedical research using AI/ML has accelerated rapidly over the last decade. Another important trend is the growth, particularly over the last 7 years, in regulatory filings and Food and Drug (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, given 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. 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. 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 ultimately have measurable positive effects on clinical outcomes.

In addition to being incomplete, Hartshorn noted, data supporting current approvals can be biased, which reduces the utility and accuracy of the clinical decision support (CDS) that AI and ML algorithms provide. Hartshorn said that such algorithms can have severe consequences if incorrectly leveraged by clinicians, raising a concern about the little to no data supporting how AI-powered CDS tools are being used by clinicians who, for example, may use them as clinical decision-making tools rather than support tools.

Nutrition for Precision Health Program

Holly Nicastro, program director in the NIH Office of Nutrition Research and program coordinator of the Nutrition for Precision Health, powered by the All of Us Research Program (NPH), 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. The program will collect data on the participants’ physiology, metabolome, microbiome, genome, dietary intake, demographics, health history, psychosocial factors, behaviors, and environment and 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 also 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.

As part of the All of Us Research Program, NPH will have access to data collected by the All of Us study and be able to study the techniques that Sazonov, Knight, and McRitchie described, such as image-assisted wearables,

random forest classification, mass spectrometry, principal component analysis, and neural networks. Nicastro further noted that there is the opportunity within NPH to merge these various data sources. For example, aligning the timing of food intake with metabolite changes may provide information on how the metabolome changes over time. Regarding DEI and accessibility, she sees a tremendous opportunity for advances in data collection to engage people who have not participated in nutrition studies and ensure that any findings are relevant to and reach diverse populations.

One of the first tasks for the program 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 predictors and the evidence behind them being included in the Dietary Guidelines for Americans, which would be a great example of precision nutrition used in a public health setting. 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, where physicians are spending less time with patients while being asked to explain ever more complicated topics such as computer-generated risk profiles. She worries that when clinicians are asked to add in more of these personalization or precision factors for a patient it will be more burdensome for the practitioner and patient and wonders whether knowing that the advice was tailored to that individual will make them better empowered to stick to their plan.

ETHICS, ACCESS, LEGAL FRAMEWORKS, AND FAIRNESS

Aaron Smith, the DeLoach professor of agricultural economics and lead of the AI Institute for Next-Generation Food Systems (AIFS) socioeconomics and ethics cluster at the University of California—Davis, discussed some of the challenges facing agriculture and food production and offered insights into how AI may provide solutions. As an economist, he sees many of these challenges in terms of externalities—that is, a side effect or consequence of an economic decision an entity makes. For example, farmers realize the benefits of fertilizing their crops but do not bear the cost of the pollution to rivers and streams. Data are a positive externality, Smith explained, as the benefits also flow beyond the provider.

Smith said that public funding of ethical technologies to benefit society is imperative when focused on solving problems arising from negative externalities and supercharging positive externalities. Regarding ethics and technology, he listed three core questions: who wins and who loses, who bears risk, and who decides. An ethical approach, he added, requires making proactive decisions to develop technologies that would benefit society rather than focusing on regulation which has limited ability to put guardrails around appropriate development and use of AI tools.

The priority, Smith said, should be investing in problems beset by externalities and develop technologies that can help surmount the challenges of human behavior, which he sees as a better path than trying to identify ways to manipulate human behavior. Examples of AI-enabled technologies relevant to agriculture include autonomous weeders, precision seeding and fertilizing, supply chain optimization, automatic detection of pathogens in food processing, diet customization tools, and engineering healthy foods.

Smith and his colleagues surveyed 1,000 Illinois corn and soybean farmers to better understand their concerns about sharing data. Nearly half of them were unconcerned, and most of those who were concerned gave multiple reasons, including others making money off their data, the top reason; an invasion of privacy; and being unsure how the data will be used. Other work his group is doing asked farmers about whether they would use AI-powered technology. Farmers voiced concerns about regulatory burdens, labor scarcity, and finance pressures. Smith and his team concluded that the main barriers to adoption are not related to trust issues but whether AI will solve real-life problems.

Janie Hipp, inaugural president and chief executive officer of Native Agriculture Financial Services, said that agricultural producers are facing unprecedented

complexity in decision making and that agricultural AI can support, but not necessarily simplify, many of those decisions. 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.

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, Hipp stated. 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 who supplies raw data? Another important question is concerns about 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.

One problem Hipp has seen as a lawyer is that farmers and ranchers sign nonnegotiable contracts to obtain the equipment and supplies they need. As a result, although most believe they own their data, they actually do not according to the contracts she has seen and analyzed. Ownership and control of farming data are significant concerns for her, and she believes they should be for policy makers, too.

Hipp said that it is hard to consider ethics in AI without confronting the nebulous concepts of community, sacredness, and protection, but not considering ethics, access, legal frameworks, and fairness in AI will lead to exploitation. It will also not allow for accessing inherent knowing and sacredness, which is a huge issue for Indigenous communities such as hers that have only just started the conversation about data ownership and AI. “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. She is concerned, too, that data and AI/ML applications are usable only if one has the time and space, and the larger players are more able to absorb data and data applications into their operations and make decisions with that data. And she noted that it will be imperative to target AI/ML applications at the level of both the larger players and vast numbers of smaller producers.

Hipp noted pitfalls and challenges associated with a failure to consider ethics, access, legal frameworks, and fairness in AI: the difficulty of recognizing and quantifying the harm caused; the rise of state-level pushback on AI 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.

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 midsize producers to have a viable path forward and stay on the land and continue producing food.

CLOSING REMARKS

To conclude the workshop, Sharon Kirkpatrick, associate professor in the School of Public Health Sciences at the University of Waterloo, summarized the workshop. Among the main points raised, she said, were that applying AI, ML, and DL to food and nutrition research involves solving a series of difficult and challenging problems, with both optimism and pessimism about success. She noted that it is important to make existing data more available, improve the representation of different populations in ongoing data collection, and address issues of data sovereignty. The field, she posited, will benefit from assembling diverse teams that collaborate and communicate effectively and develop a shared vocabulary and culture and by engaging communities and other stakeholders early in study design. Kirkpatrick continued by stating that

the development and use of AI/ML tools be done in a manner that supports equity, fairness, and justice for all end users and reflects the broader context of social and structural determinants of health and nutrition. AI is useful, she observed, but it is not magic and requires human expertise to be most useful.

Kirkpatrick noted the general lack of institutional support for team-based science and wondered how to address this problem. Mejía-Argueta acknowledged that this is a difficult situation and suggested ways to bolster team performance:

• Identifying the key performance indicators for each member of a team and similarities across the team that can be leveraged to benefit everyone regarding institutional recognition, perhaps by enabling different members to serve as the primary authors on publications reporting the findings.
• Using grant funds to collaborate with colleagues from countries outside of the United States that face similar problems and challenges.
• Inviting people from industry or nonprofit organizations to provide feedback on programmatic activities to expose team members to the need to ensure that their work is useful.
• Gaining the trust of others to create alliances that are sustainable over the long term and that can create their own key performance indicators.
• Developing a culture that creates synergies and ensures the team is robust and able to address the same problem from different perspectives.

Das said that not everyone would need to be equipped for assembling large teams but could be ready to contribute to efforts that would benefit from their domain expertise and that this will require keeping up with and understanding the landscape and comprehensive nature of ongoing research and providing expertise within that context. Odoms-Young noted the movement at some institutions to change the main performance indicators used for promotion and tenure decisions to reflect the value of team-based research or community-based participatory research, for example. The key is having the right metrics of success in place that perhaps professional organizations could develop to reflect new ways of engaging in research.

Smith said that deep disciplinary expertise is needed for the field to realize its potential. Therefore, it is important, he said, to ensure that training does not dilute that, by only producing people who are good at everything; generalists help disciplinary experts talk to one another.

DISCLAIMER This Proceedings of a Workshop—in Brief has been prepared by Joe Alper as a factual summary of what occurred at the meeting. The statements made are those of the rapporteur or individual workshop participants and do not necessarily represent the views of all workshop participants; the planning committee; or the National Academies of Sciences, Engineering, and Medicine.

*The National Academies of Sciences, Engineering, and Medicine’s planning committees are solely responsible for organizing the workshop, identifying topics, and choosing speakers. The responsibility for the published Proceedings of a Workshop—in Brief rests with the institution. The Planning Committee comprises Rodolphe Barrangou (cochair), North Carolina State University; Sharon I. Kirkpatrick (cochair), University of Waterloo; Becca Jablonski, Colorado State University; Anant Madabhushi, Georgia Institute of Technology and Emory University; Carmen D. Tekwe, Indiana University at Bloomington; and Diana Thomas, U.S. Military Academy at West Point.

REVIEWERS To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by Marie E. Latulippe, Institute for the Advancement of Food and Nutrition Science; Nicholas J. Schork, The Translational Genomics Research Institute, part of City of Hope; Muhammed Y. Idris, Morehouse School of Medicine; Holly Nicastro, National Institutes of Health; and Whitney R. Robinson, Duke University School of Medicine. Leslie Sim, National Academies of Sciences, Engineering, and Medicine, served as the review coordinator.

SPONSORS This workshop was supported by Texas A&M University (AB0767199) and the U.S. Department of Agriculture Agricultural Research Service (59-0204-0-003).

STAFF Alice Vorosmarti, Melanie Arthur, and Ann L. Yaktine, Food and Nutrition Board, Health and Medicine Division, National Academies of Sciences, Engineering, and Medicine.

For additional information regarding the workshop, visit https://www.nationalacademies.org/event/40460_10-2023_the-role-of-advanced-computation-predictive-technologies-and-big-data-analytics-related-to-food-and-nutritionresearch-a-workshop.

Suggested citation: National Academies of Sciences, Engineering, and Medicine. 2024. The role of advanced computation, predictive technologies, and big data analytics in food and nutrition research: Proceedings of a workshop—in brief. Washington, DC: National Academies Press. https://doi.org/10.17226/27482.

Health and Medicine Division Copyright 2024 by the National Academy of Sciences. All rights reserved.