NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines
New hierarchical AI framework outperforms GPT-4.1, delivering personalized nutrition plans for patients with multiple chronic conditions.
A research team from Emory University and Georgia Tech has introduced NutriOrion, a novel hierarchical multi-agent AI framework designed to solve the complex challenge of creating personalized nutrition plans for patients with multiple chronic conditions (multimorbidity). The system addresses a critical limitation of single large language models (LLMs) like GPT-4.1, which often suffer from 'context overload' and 'attention dilution' when processing high-dimensional patient profiles involving various diseases, medications, and dietary guidelines.
Technically, NutriOrion employs a 'parallel-then-sequential' reasoning topology. It first decomposes the nutrition planning task across specialized domain agents, each working with isolated contexts to prevent anchoring bias. This is followed by a conditional refinement stage. A core innovation is its multi-objective prioritization algorithm that resolves conflicting dietary requirements and a safety constraint mechanism. This mechanism injects pharmacological contraindications as hard negative constraints during plan synthesis, ensuring clinical validity 'by construction' rather than through less reliable post-hoc filtering. For healthcare interoperability, the framework outputs plans formatted to the ADIME (Assessment, Diagnosis, Intervention, Monitoring, Evaluation) standard and FHIR R4 resources.
In evaluation on 330 real stroke patients with multimorbidity, NutriOrion outperformed several baselines, including GPT-4.1 and other multi-agent architectures. Key results include a drug-food interaction violation rate of just 12.1%, and strong evidence of personalization, shown by negative correlations (-0.26 to -0.35) between patient biomarkers and recommended risk nutrients. The AI-driven plans led to clinically meaningful dietary improvements: a 167% increase in fiber, a 27% increase in potassium, alongside reductions of 9% in sodium and 12% in sugars. This demonstrates a significant step toward automated, scalable, and safe clinical decision support for personalized nutrition.
- Outperforms GPT-4.1 with a 12.1% drug-food interaction violation rate in tests on 330 stroke patients.
- Uses a 'parallel-then-sequential' multi-agent architecture to prevent LLM context overload and anchoring bias.
- Delivered a 167% increase in fiber and 27% more potassium in personalized plans while cutting sodium and sugars.
Why It Matters
Automates creation of safe, personalized diet plans for complex patients at scale, directly integrating into clinical workflows.