What is this case study about?

This case study documents the development, validation, and implementation of PlantVillage FRANI (Food Recognition AI for Nutritional Intelligence), an AI-powered mobile application for dietary assessment and behavioral nudging. Led by Aulo Gelli (Senior Research Fellow at the International Food Policy Research Institute, IFPRI), the project is a collaboration with Penn State University's PlantVillage, the University of Ghana, and the National Institute of Nutrition in Vietnam.

PlantVillage FRANI addresses a critical gap in nutrition evaluation: traditional dietary surveys are expensive, complex, and difficult to scale. By using computer vision to analyze smartphone photos of meals, the app provides accurate nutrient estimates at a fraction of the cost of traditional methods—while also enabling real-time behavioral nudging to promote healthier dietary choices.

Why was this approach necessary?

The challenge of unhealthy diets

Unhealthy diets are a major global health crisis. According to the Global Burden of Disease analyses (Afshin et al., 2019), unhealthy diets are a major risk factor for global mortality and morbidity, and improvements in diet could potentially prevent one in every five deaths globally. Risks are exacerbated by urbanization and the "nutrition transition"—increased consumption of unhealthy processed foods and decreased physical activity, leading to rising rates of overweight and obesity (Popkin et al., 2020).

The adolescent nutrition gap

Nutrition and diets during school age and adolescence are critical for development and health. For adolescent girls, they also affect the survival and wellbeing of their children. Yet school age and adolescent nutrition and diets have been largely overlooked (Patton et al., 2021).

The data gap

Collection and use of dietary data is costly and complex (Bell et al., 2017). Traditional dietary surveys commonly use the multi-pass 24-hour recall (24HR) method, which has been validated for use in LMICs. However:

• High cost: Undertaking 24HR costs approximately $500 per recall (Adams et al., 2022).
• Complexity: The age at which children and adolescents can accurately self-report food intake without support is unclear, with varying respondent-related challenges (Livingstone & Robson, 2000).
• Limited technology adoption: Technology-assisted dietary assessment tools have been proposed but have lacked feasibility and validity assessments in LMICs (Bell et al., 2017).

What is PlantVillage FRANI?

PlantVillage FRANI is an AI-based mobile phone application for dietary assessment that uses smartphone photos with computer vision technology to identify foods and estimate portion sizes from plate images. The app produces nutrient estimates comparable to weighed records and on par with 24-hour recalls by trained dietitians—at a fraction of the cost.

Key features

• Semantic segmentation: The app uses pixel-wise classification to identify individual foods in images, trained using TensorFlow over 3-4 days.
• Portrait mode: Enhances food recognition by focusing on the meal while blurring background distractions.
• Offline capability: Enables use in low-connectivity settings common in LMICs.
• Gamified nudging: Includes a version with gamified nudging based on USDA's "Start Simple with MyPlate" application to encourage healthier choices.
• Real-time dashboard: Provides real-time visualizations for program administrators to monitor meal delivery and nutritional quality.

Development process

The development of PlantVillage FRANI followed a user-centered, iterative design process:

Training datasets and model architecture

Semantic segmentation model

The food recognition model uses semantic segmentation with a convolutional encoder-decoder architecture, enabling pixel-wise classification of foods in images. Training takes approximately 3-4 days using TensorFlow.

Training datasets

Beyond segmentation: Integrating LLMs

The team is now integrating Large Language Models (LLMs) into the food recognition pipeline to increase flexibility and reduce training costs. This advancement is currently being validated in the USA.

Validation results: Accuracy and cost efficiency

Validation studies

Validation studies in Vietnam and Ghana demonstrated that PlantVillage FRANI accurately assesses nutrient intake and is at least as accurate as 24-hour recalls conducted by trained dietitians in school-age children and adolescents (Diop et al., 2025).

Cost comparison

This represents a 1,000x reduction in cost, enabling high-frequency monitoring previously not feasible.

Behavioral nudging: Can FRANI improve diets?

The team tested whether FRANI could be used not only to assess diets but also to improve them through gamified nudging, based on Food Based Dietary Guidelines and the USDA "Start Simple with MyPlate" framework.

Vietnam pilot: EAT-Lancet diet score improvement

A randomized pilot tracked diets of 36 adolescent females (12-19 years) over 30 days using FRANI (with nudging) and FRANI Control (without nudging).

Result: Increase in mean daily EAT-Lancet diet score of ~1.1 points over a base level of 3.7 (p=0.032). (Source: Braga et al., 2023)

Ghana pilot: Dietary Diversity Score improvement

A randomized pilot tracked diets of 60 youth (18-24 years) over 55 days using FRANI and FRANI Control.

Result: Participants in the treatment group had higher Dietary Diversity Scores (DDS) compared to the control group during the pilot period (B: 1.07; 95% CI: 1.00, 1.14).

Real-time monitoring dashboard

Beyond individual dietary assessment, FRANI includes a dashboard that provides real-time visualizations for program administrators, enabling them to quickly identify gaps in school meal programming:

• Meal delivery tracking: Identify schools not regularly providing meals
• Coverage monitoring: Track dates when meals are not being provided across the district
• Nutritional quality: Monitor challenges in meeting school meal nutritional standards
• Food-level analysis: Examine specific foods and portion sizes in real-time

Key takeaways

Main reference & original sources

📘 This case study is based on the presentation by Aulo Gelli and the PlantVillage FRANI team, as documented in:
1. WFP Evaluation (2026, March 23). Four examples of how AI, machine learning and data innovation are reshaping evaluation and research. Medium. https://wfp-evaluation.medium.com/four-examples-of-how-ai-machine-learning-and-data-innovation-are-reshaping-evaluation-and-research-f4d1b3cf4e74
2. Gelli, A. (2025, December). Nudging for Good: AI-Driven Diagnostics and Behaviour Change to Improve Diets and Nutrition. Presentation slides. WFP Global Impact Evaluation Forum 2025. Download full presentation (PDF)

Resources for further learning

• WFP Evaluation – Medium Blog
• EvalCommunity – AI in M&E Course
• Download: Nudging for Good (PDF)
• PlantVillage – Penn State University
• EvalCommunity Services & Resources