Abigail Horn, lead scientist at USC Viterbi’s Information Sciences Institute (Photo/Courtesy of Abigail Horn)
Choosing somewhere to eat is no easy task. There are a seemingly endless number of options with differences in cuisine, ambiance, and so much more. And after hours of decision making, it’s usually easier to eat at a place that is familiar.
Unfortunately, the thought of nutrition often goes by the wayside. Due to a lack of information, people all-too-often make unhealthy food choices in their neighborhoods. Whether in a food desert, an area with limited access to fresh, healthy food, or a food swamp, an area with an overabundance of fast-food, or somewhere in between, Abigail Horn and her team believe she has a solution that would make it easier for people to eat better when they go to eat, and likewise, for policy decisionmakers to assess disparities in access to healthy food.
Horn, a lead scientist at USC Viterbi’s Information Sciences Institute (ISI) and research assistant professor at the Daniel J. Epstein Department of Industrial and Systems Engineering, has found that predicting the nutrient density of meal options from information accessible on menus can help to identify healthy food options in a given area, which, she believes, could contribute to improving public health. This is in the context of extensive research efforts, including her own, that have shown that people’s food environments have a large effect on their diets and risk for diet-related disease, such as obesity, diabetes, cancer, and heart disease.
Horn and her team, fellow ISI researcher Keith Burghardt and Alex DongHyeon Seo, a recent graduate of the masters of applied data science program, created MINT: Menu Item to NuTrient. Burghardt, a computer scientist, led the development of MINT, a large-scale AI-based model that can predict the nutritional quality of a restaurant by using the information available on menus.
Trained on large amounts of meal and recipe data and applied to menu data from about 600,000 restaurants across the U.S., MINT can calculate the nutrient density of foods based on their names alone. Although not perfect, restaurant-specific foods like “Volcano Nachos” or a “Buddha Bowl” can be analyzed from the structure of language learned from previous data. MINT predicts each menu item’s nutritional quality. The system then compiles all items for a given restaurant into a single score called the Restaurant Nutrient Density (RND); a lower number identifying unhealthy options, like Shake Shack at 0.30, and a higher number identifying healthier options, like Sweetgreen at 1.33.
“I think we have this unprecedented perspective into the healthiness of restaurants across the country, where they’re located, and the populations that have access to them,” Horn said. “You can use this information to look for neighborhoods with healthier restaurants.”
More than making it easier for restaurant-goers to make healthier choices, the model was developed to support policymakers. For example, in Los Angeles there is a program called the Restaurant Meals Program, which gives people that are at a high risk for chronic hunger the option to use CalFresh benefits to buy prepared meals at approved participating restaurants. Many of those approved restaurants are fast food places serving primarily low-nutrient, energy-dense foods. The list of restaurants is pages long, organized strictly by zip code. But by leveraging MINT, state governments could highlight the healthiest restaurants in an area.
“If we’re making these big decisions about where to make changes to neighborhoods, make changes to food environments to make them healthier, and meeting people where their needs are, this gives you a lot more fine-grained information to do that,” Horn said. “You could think about resource allocation and policy investment, really taking in this information as an important determinant to those kinds of decisions.”
Horn said she and her team learned a lot about the communities they investigated through statistical modeling. For example, the researchers discovered less healthy food options were more common in low-income neighborhoods. As well, there were ethnic disparities. Neighborhoods with predominantly Black residents tended to be more associated with lower nutritional quality, while the opposite was true for neighborhoods that were predominantly Asian or Hispanic, oftentimes determined by traditional ingredients favored in these cultures.
“There are lots of beans, high fiber, and vegetables used in these restaurants, for example all salsas have vegetables,” she said. “And there’s a lot of grilled protein sources. They hit higher protein and higher fiber than a lot of the fast food or other Western-normative style restaurants.”
Eating out is as popular as ever, particularly with the rise of food delivery services like Uber Eats and DoorDash. Over 50 percent of people’s food budgets are spent at restaurants or ordering food in, according to Horn, making her research more relevant than ever.
But she and her team aren’t stopping with just nutritional information.
“There’s definitely more work that can be done. I’d like to be able to predict other features of the food, beyond the nutrients,” said Horn. “We want to do a better job of predicting the ingredients that are used, the amount that they’re used in, the degree of processing, and what we call the food matrix: if something is blended or processed versus if you’re eating it with a fork and knife.”
And these goals can be reached.
In the future, their work has the potential to be further supported and sophisticated through collaborations. It dovetails perfectly with USC President Carol Folt’s Frontiers of Computing “moonshot,” a 10-year initiative to accelerate research and innovation in advanced computing and its related applications with a focus on ethical considerations.
Published on September 3rd, 2024
Last updated on October 18th, 2024
