PanaCea: Clinical Hypergraph Framework for Health-Aware Personalized Diet Recommendation

Junwei Wu, Runze Yan, Yuzhang Xie, Zidu Xu, Jiaying Lu, Hanqi Luo, Xiao Hu, Carl Yang

Keywords: Nutrition Informatics; Health-Aware Recommendation; Hypergraph Learning; Multimodal Fusion

Abstract: Diet quality plays a critical role in health outcomes, making personalized food recommendations vital to encourage healthier eating. Effective diet recommendations must jointly consider user profiles, eating history, health conditions, and food nutritional and quality data. Current methods are flawed, often focusing on limited factors, suggesting unhealthy or unpalatable foods, or proposing micronutrient-rich foods unsuitable for certain health conditions. The core challenge is managing the heterogeneous, hierarchical, and complex interconnected nature of dietary data, including diverse profiles and meals with intricate nutritional interactions beyond the capability of previous approaches. We introduce PanaCea, a hypergraph neural network framework that integrates authoritative nutrient composition databases, large-scale population dietary intake data, electronic health records, and food healthfulness scoring systems to provide personalized and health-aware food recommendations. Evaluation in NHANES, a nationally representative dietary survey, demonstrates our framework's ability to effectively balance recommendation relevance with nutritional safety, achieving superior performance on ranking metrics and novel nutrition-focused measures. PanaCea provides a comprehensive framework for clinically informed dietary recommendations that can be adapted to various health conditions and datasets, offering practical potential to support healthier eating behaviors across diverse populations while maintaining individual personalization.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 2740