Human-AI Collaboration for Allergen-free Diet Generation

Abstract

Balanced allergen-free diet planning is essential for the management of food allergy, which should simultaneously satisfy allergen removal and nutritional adequacy while preserving dietary composition. However, this task is challenging even for expert dietitians. The diet planning itself has the complexity of finding feasible diets among menu combinations, where previous studies considered it as a combina- torial optimization problem with constraints on nutritional requirements. However, these studies were not able to satisfy the dietary composition that comes with implicit, dietitian-dependent constraints. Under the scarcity of high-quality data, training implicit patterns comes with a second challenge. Furthermore, food allergy often involves with combinations of multiple allergens, which also leads to the computational challenge. Consequently, these challenges make it difficult to plan balanced allergen-free diets even for expert dietitians. To address these challenges, this study highlights the need for an assist artificial intelligence (AI) system that could support dietitians’ allergen-free diet planning task more ef- ficiently. Therefore, I propose two complementary solutions to improve dietitians’ workflow efficiency with developed methods. I first introduced a framework for nutritionally adequate allergen-free diet planning, which is an embedding-guided post-hoc editing framework. High-quality data is created through a human-in-the- loop process to satisfy the nutritional requirements and dietary compositions. Then, the food embedding is trained with nutrition-ingredient information and dietary composition to provide substitutions for the allergen-containing menus. This framework has been clinically validated on real-world pediatric food allergy cohorts in a human-AI collaboration manner. The results show that generated diets, increased caregivers’ personalized self-efficiency in dietary management, achieved high satisfaction scores, and increased efficiency of dietitians’ workflow by reducing the editing effort. The second proposed method, attribute-aware controllable menu editor (ACMEditor) for allergen free diet planning, is a one-shot, ingredient-aware controllable diet generator that conditions on multi- allergen sets with the attribute co-attention layer. The three-stage training process with reinforcement learning and graph embedding-guided decoding strategy allows the model to encourage the simultaneous satisfaction for both allergen removal and dietary composition together. Across quantitative metrics and human evaluation, ACMEditor achieved perfect allergen removal, stronger composition measures than baselines. In the experiment analysis of two scenarios of human-AI collaboration, results show a statis- tical improvement on dietitians’ workflow in both composition-editing and AI-assistant scenario. These results indicate that the two proposed methods demonstrated the ability to assist dietitians’ allergen-free diet planning task. From the proposed studies, I addressed real-world challenges in allergen-free diet planning and showed that an AI-assistant system can effectively support tasks that dietitians typically find burden- some, enabling practical use in pediatric care.