(P03-06-26) Leveraging AI for Dietary Assessment – Development and Validation of the PlantVillage Food App for U.S. College Students

Objectives: The PlantVillage Food (PV Food) mobile phone app uses AI to estimate nutrients from photographs of food and encourages healthy choices. The objective of the study is to adapt and validate the PV Food app in a U.S. college population.

Methods: The PV Food app was adapted for the US by integrating a new large language model (LLM) technology for food recognition and portion estimation and modifying the user interface for use on iOS devices. Foundational work included developing a PV Food and Nutrient Database by integrating data on standard portions from USDA FoodData Central, food composition information from the 2021-2023 USDA Food and Nutrient Database for Dietary Studies, and recipe information from the University of California, Davis Dining Commons (DC). The validation study followed a cross-sectional, repeated-measures design. Eligible participants were 18-24 years with a residential meal plan and an iOS device; those with a history of/at risk for eating disorders were excluded. Research assistants conducted weighed food records (WFR) and participants (UC Davis college students) used the PV Food app to photograph all foods consumed in a DC setting on two nonconsecutive days between Nov-Dec 2025. Equivalence was assessed by comparing mean intake ratios (PV Food/WFR) within 20% error margins, using mixed-effects regression models adjusted for repeated measures. Concordance correlation coefficient (CCC) was used to examine agreement between methods.

Results: All study procedures were completed by 66 college students. The mean age was 18.5±1.0 years. Overall, 62% had a normal BMI, more than 50% self-identified as Hispanic, and nearly 60% were first-generation college students. The mean±SD total daily energy intake consumed in the DC and reported in the WFR was 741±446 kcal, of which 41.3±15.8% came from carbohydrates, 20.5±7.0% from protein, and 38.3±12.1% from fat. Overall, 98.6% of foods were correctly recognized by the app. CCC varied between 0.60-0.80 depending on the nutrient considered.

Conclusions: Although improvements in the LLM classification and portion estimation are underway, emerging evidence in this innovative research suggests that this technology has potential for accurate dietary assessment with U.S. college students.

Funding Sources: American Heart Association & UC Davis Senate Grant