Paper deep dive
NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines
Junwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, Carl Yang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/20/2026, 9:37:57 PM
Summary
The paper introduces NutriOrion, a hierarchical multi-agent framework for personalized nutrition intervention for patients with multimorbidity. It addresses the limitations of single-agent LLMs by using a parallel-then-sequential reasoning topology with specialized domain agents (Body, Clinical, Medication, Diet) to mitigate context overload and anchoring bias. The system integrates clinical guidelines (DASH, ADA) and pharmacological data (DailyMed) via Retrieval-Augmented Generation (RAG) to ensure clinical validity. A key feature is the Safety Constraint Mechanism, which injects drug-nutrient interaction contraindications as hard negative constraints during synthesis. The framework outputs structured ADIME plans mapped to FHIR R4 resources. Evaluated on 330 stroke patients, NutriOrion outperforms baselines like GPT-4.1, achieving a 12.1% drug-food interaction violation rate and significant improvements in dietary metrics such as fiber and potassium intake.
Entities (10)
Relation Signals (10)
NutriOrion → achievesmetric → 12.1% drug-food interaction violation rate
confidence 95% · It achieves a 12.1 percent drug-food interaction violation rate
NutriOrion → improvesmetric → 167% increase in fiber
confidence 95% · yields clinically meaningful dietary improvements, including a 167 percent increase in fiber
NutriOrion → mapsto → ADIME
confidence 95% · NutriOrion maps synthesized insights into the ADIME standard
NutriOrion → mapsto → FHIR R4
confidence 95% · NutriOrion maps synthesized insights into the ADIME standard and FHIR R4 resources.
NutriOrion → usestopology → Parallel-then-Sequential
confidence 95% · NutriOrion, a hierarchical multi-agent framework featuring a Parallel-then-Sequential reasoning topology.
NutriOrion → evaluatedon → NHANES
confidence 90% · evaluation on 330 multimorbid stroke patients from the the National Health and Nutrition Examination Survey (NHANES) dataset
NutriOrion → outperforms → GPT-4.1
confidence 90% · NutriOrion significantly outperforms multiple baselines, including GPT-4.1
NutriOrion → usesretrievalsource → ADA
confidence 90% · dual-source retrieval system that accesses both clinical guidelines (DASH, ADA)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Personalized nutrition intervention for patients with multimorbidity is critical for improving health outcomes, yet remains challenging because it requires the simultaneous integration of heterogeneous clinical conditions, medications, and dietary guidelines. Single-agent large language models (LLMs) often suffer from context overload and attention dilution when processing such high-dimensional patient profiles. We introduce NutriOrion, a hierarchical multi-agent framework with a parallel-then-sequential reasoning topology. NutriOrion decomposes nutrition planning into specialized domain agents with isolated contexts to mitigate anchoring bias, followed by a conditional refinement stage. The framework includes a multi-objective prioritization algorithm to resolve conflicting dietary requirements and a safety constraint mechanism that injects pharmacological contraindications as hard negative constraints during synthesis, ensuring clinical validity by construction rather than post-hoc filtering. For clinical interoperability, NutriOrion maps synthesized insights into the ADIME standard and FHIR R4 resources. Evaluated on 330 stroke patients with multimorbidity, NutriOrion outperforms multiple baselines, including GPT-4.1 and alternative multi-agent architectures. It achieves a 12.1 percent drug-food interaction violation rate, demonstrates strong personalization with negative correlations (-0.26 to -0.35) between patient biomarkers and recommended risk nutrients, and yields clinically meaningful dietary improvements, including a 167 percent increase in fiber and a 27 percent increase in potassium, alongside reductions in sodium (9 percent) and sugars (12 percent).
Tags
Links
- Source: https://arxiv.org/abs/2602.18650v1
- Canonical: https://arxiv.org/abs/2602.18650v1
Trouble viewing inline? Open PDF directly →
Full Text
71,688 characters extracted from source content.
Expand or collapse full text
NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Junwei Wu junwei.wu@emory.edu Emory University Atlanta, GA, USA Runze Yan runze.yan@emory.edu Emory University Atlanta, GA, USA Hanqi Luo hanqi.luo@emory.edu Emory University Atlanta, GA, USA Darren Liu darren.liu@emory.edu Emory University Atlanta, GA, USA Minxiao Wang minxiao.wang@emory.edu Emory University Atlanta, GA, USA Kimberly L. Townsend ktownsend3@twu.edu Texas Woman’s University Denton, TX, USA Lydia S. Hartwig lrschinn@gmail.com Cecelia Health New York, NY, USA Derek Milketinas dmiketinas@twu.edu Texas Woman’s University Denton, TX, USA Xiao Hu xiao.hu@emory.edu Emory University Atlanta, GA, USA Carl Yang j.carlyang@emory.edu Emory University Atlanta, GA, USA Abstract Personalized nutrition intervention for patients with multimorbid- ity is critical for improving health outcomes, yet remains challeng- ing as it requires the simultaneous integration of heterogeneous clinical conditions, medications, and dietary guidelines. Single- agent Large Language Models (LLMs) often suffer from context over- load and attention dilution when processing such high-dimensional patient profiles. We introduce NutriOrion, a hierarchical multi- agent framework featuring a Parallel-then-Sequential reasoning topology. NutriOrion decomposes nutrition planning into special- ized domain agents with isolated contexts to mitigate anchoring bias, followed by a conditional refinement stage. The framework incorporates a multi-objective prioritization algorithm to resolve conflicting dietary requirements and a Safety Constraint Mecha- nism that injects pharmacological contraindications as hard neg- ative constraints during synthesis, ensuring clinical validity by construction rather than post-hoc filtering. To ensure clinical inter- operability, NutriOrion maps synthesized insights into the ADIME standard and FHIR R4 resources. Evaluated on 330 stroke patients Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Conference’17, Washington, DC, USA © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n with multimorbidity, NutriOrion significantly outperforms multi- ple baselines, including GPT-4.1 and alternative multi-agent archi- tectures. It achieves a 12.1% drug-food interaction violation rate, demonstrates strong personalization with negative correlations (−0.26 to−0.35) between patient biomarkers and recommended risk nutrients, and produces clinically meaningful improvements, including a 167% increase in fiber and 27% increase in potassium, alongside reductions in sodium (9%) and sugars (12%). CCS Concepts • Computing methodologies→Multi-agent systems;• Ap- plied computing→Health informatics;• Information sys- tems→ Information retrieval. Keywords Multi-agent Systems, Information Retrieval, Health Informatics, Decision Support Systems ACM Reference Format: Junwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang. 2026. NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines. In . ACM, New York, NY, USA, 14 pages. https://doi.org/10.1145/n.n 1 Introduction Dietary intervention is critical for managing chronic diseases, which account for over 70% of global mortality [12]. While structured nutri- tion can achieve therapeutic effects comparable to pharmacological treatments [14,17], personalization is hindered by multimorbid- ity, affecting 40% of chronic disease patients, where co-occurring arXiv:2602.18650v1 [cs.MA] 20 Feb 2026 Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang Figure 1: Conceptual comparison of reasoning paradigms for high-stakes decision making. (a) Monolithic LLM Approach: Standard single-agent models process heterogeneous data in a unified context window. This often leads to information entanglement, where conflicting constraints are hallucinated or ignored due to attention dilution. (b) NutriOrion (Ours): A hierarchical modular framework that decouples reasoning into specialized streams. By isolating domain contexts and enforcing deterministic safety gates, NutriOrion synthesizes complex inputs into precise, verifiable, and structured interventions without reasoning collapse. conditions often demand conflicting dietary strategies [10,35]. Fur- thermore, a severe global shortage of nutrition professionals (1:5,000 ratio) makes manual integration of patient data with complex clini- cal guidelines nearly impossible at scale [18, 38]. Recent years have witnessed growing research in computational dietary recommendation systems, with several distinct method- ological approaches emerging [27,42]. Early systems employed collaborative filtering and content-based methods, recommending foods by matching user preferences with similar users or item attributes [3,13,20]. A second line of work developed constraint- based approaches that explicitly encode dietary restrictions as hard rules, filtering out foods that violate specific nutritional constraints such as sodium limits for hypertension management [20,32]. In contrast, optimization-based methods formulate meal planning as mathematical programming problems, where algorithms search for meal combinations that simultaneously optimize multiple nu- tritional objectives under various constraints [16,28]. More re- cent hybrid approaches combine rule-based medical knowledge with machine learning models, using expert-defined nutritional guidelines to constrain the recommendation space while allowing data-driven learning to adapt to individual user preferences and dietary patterns [1,29]. The latest deep learning methods learn representations by embedding foods and users into continuous vector spaces, enabling prediction of health outcomes and personal- ized recommendations based on individual profiles and contextual factors [26, 33]. However, despite these methodological advances in computa- tional dietary recommendation, existing systems remain inadequate for clinical deployment due to four fundamental limitations. First, they optimize for generic health scores rather than disease-specific clinical objectives defined in existing dietary guidelines. Most sys- tems use composite wellness metrics or broad nutritional balance scores without operationalizing the precise quantitative targets specified in evidence-based protocols for managing specific chronic diseases. Second, they fail to provide personalized modifications to actual dietary intake. Instead of analyzing what patients cur- rently eat and suggesting actionable adjustments, most systems generate idealized meal plans disconnected from real-world eating behaviors, making adherence difficult. Third, they cannot handle multimorbidity scenarios with conflicting dietary requirements. Existing approaches either process diseases independently (gen- erating incompatible recommendations) or aggregate objectives into a single function (losing clinical prioritization logic). Fourth, they lack robust verification of medication-diet interactions and strict adherence to existing dietary guidelines for each comorbid- ity. With multiple medications and diseases, existing systems do not perform systematic cross-domain verification for safety across pharmacological constraints [4, 30]. Large Language Models (LLMs) offer new possibilities for ad- dressing these challenges through their capabilities in multi-domain reasoning and integrating heterogeneous information [36,49]. How- ever, existing work shows that monolithic LLM architectures suffer from reasoning collapse when processing high-dimensional feature spaces [24,31]. In personalized nutrition for multimorbidity, single- agent models face this challenge when processing comprehensive patient profiles containing dozens of heterogeneous features: demo- graphics, anthropometrics, laboratory biomarkers, medication lists, dietary histories, and existing dietary guidelines, as illustrated in Figure 1. As context complexity increases, models generate recom- mendations violating safety constraints, hallucinate interventions NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA contradicting patient data, and exhibit attention dilution where important constraints get overridden [5,34]. This occurs because monolithic architectures process all information through unified attention, causing cross-domain interference when reasoning spans multiple knowledge domains [7,24]. Furthermore, probabilistic safety mechanisms (RLHF, constitutional AI) cannot provide deter- ministic guarantees required for medical decision support where rare failures cause serious adverse events [5, 34]. We propose NutriOrion, a hierarchical multi-agent framework that modularizes nutrition intervention into a Parallel-then-Sequential reasoning topology. In Stage 1, domain-specific agents (Body, Clin- ical, Medication, and Diet) maintain isolated contexts to perform independent assessment grounded in dual-source retrieval from clinical guidelines and pharmacological databases [22]. This in- dependence mitigates anchoring bias and prevents cross-domain hallucination. In Stage 2, specialized dietitians perform conditional refinement, using diagnostic insights from Stage 1 as reasoning constraints. To resolve multimorbidity conflicts, a Health Priori- tization Agent performs synthesis using multi-objective scoring 푆(푗)based on severity, urgency, and modifiability. Crucially, rather than relying on post-hoc filtering, NutriOrion implements a Safety Constraint Mechanism that extracts pharmacological contraindi- cations and injects them as hard negative constraints during final generation, ensuring clinical validity by construction. Finally, the framework projects synthesized insights into the ADIME standard [39] and performs an isomorphic mappingΦto FHIR R4 resources, ensuring seamless EHR interoperability. Contributions. This work makes four primary contributions: (1) a hierarchical multi-agent architecture with isolated contexts and domain-specific RAG that prevents reasoning collapse, grounds recommendations in existing dietary guidelines for disease-specific clinical objectives, and enables personalized modifications to ac- tual dietary intake; (2) a Severity-Urgency-Modifiability scoring mechanism enabling dynamic, patient-specific guideline conflict resolution in multimorbidity scenarios; (3) a hybrid architecture separating probabilistic generation from deterministic verification against medication-diet contraindication databases, providing low- failure-rate safety guarantees; and (4) comprehensive evaluation on 330 multimorbid stroke patients from the the National Health and Nutrition Examination Survey (NHANES) dataset [19] across five complementary dimensions: output structure and actionability, medication safety, clinical appropriateness, personalization, and food quality, demonstrating that NutriOrion significantly outper- forms strong baselines including GPT-4.1 and Claude-Sonnet-4 with various prompting strategies and alternative multi-agent architec- tures. 2 Related Work 2.1 LLM Agents in Healthcare Large Language Models (LLMs) have demonstrated significant po- tential in clinical tasks, ranging from medical question-answering [37] to generating clinical notes [50]. Recent work has explored deploying multi-agent systems to mimic clinical workflows and im- prove reasoning capabilities [40,44]. For instance, MedAgents [40] utilizes a collaborative multi-agent structure to enhance diagnosis accuracy. However, existing approaches often rely on general medi- cal knowledge and lack specific mechanisms for handling complex, multi-modal patient data (NHANES) and resolving multi-morbidity conflicts in nutrition intervention. NutriOrion addresses this by introducing a specialized parallel-then-sequential topology with explicit bias mitigation and constraint propagation mechanisms, designed specifically for personalized nutrition planning. 2.2Retrieval-Augmented Grounding and Safety Retrieval-Augmented Generation (RAG) has emerged as a critical technique to mitigate LLM hallucination by grounding generation in external knowledge bases [15,22]. In the medical domain, RAG has been applied to ground responses in verified clinical guide- lines and biomedical literature [9,52]. While effective for general information retrieval, prior work often lacks robust mechanisms to enforce strict clinical safety constraints, particularly regarding drug-nutrient interactions (DNIs). NutriOrion advances this by integrating a dual-source retrieval system that accesses both clini- cal guidelines (DASH, ADA) and structured pharmacological data (DailyMed). Crucially, we implement a retrieval-augmented safety constraint mechanism that treats retrieved DNI information as hard negative constraints during synthesis, ensuring output clinical validity by construction. 3 Methodology We formulate the problem of personalized nutrition intervention as a structured inference task mapping a multi-modal patient state푋 to a clinically valid action plan푌. To achieve this, we propose Nutri- Orion, a framework shown in Figure 2 that integrates hierarchical agentic reasoning with strict evidence grounding. 3.1 Framework Formulation LetPdenote the patient population. For each patient푝 ∈ P, the input state푋 푝 ∈Xis a composite vector of clinical and demographic features: 푋 푝 =[퐴 푝 ,퐵 푝 ,푀 푝 ,퐷 푝 ,푆 푝 ](1) where퐴 푝 represents anthropometrics (e.g., BMI),퐵 푝 clinical biomark- ers (e.g., HbA1c, BP),푀 푝 the medication profile,퐷 푝 dietary history, and 푆 푝 sociodemographic factors. The objective is to learn a mapping function퐹:X → Ythat generates an output푌 푝 strictly adhering to the ADIME (Assessment, Diagnosis, Intervention, Monitoring, Evaluation) clinical standard. We decompose퐹into a hierarchical composition of specialized agent functions 푓 푖 , grounded by a retrieval functionR(·): 푌 푝 = 퐹(푋 푝 ;K,Θ)=O◦H(푋 푝 ,R(푋 푝 ,K))(2) Here,Krepresents the external knowledge base,His the hierarchi- cal reasoning engine, andOis the output synthesis function. The system parametersΘ(LLM weights) remain fixed with temperature 푇= 0 to ensure deterministic execution. 3.2 Retrieval-Augmented Grounding (R) To ensure clinical validity and mitigate hallucination, we construct a dual-source retrieval systemRthat maps agent queries푞to evi- dence contexts퐶. 3.2.1 Guideline Knowledge Manifold. We define a vector space V ⊂R 푑 populated by embeddings of verified clinical guidelines (DASH, ADA, USDA). Let퐷 푔푢푖푑푒 be the set of guideline documents. Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang Figure 2: The NutriOrion System Pipeline. The framework integrates three modular components: (Left) Retrieval-Augmented Grounding (R): Ingestion of patient profiles and vectorization of clinical guidelines (e.g., DASH, ADA); (Center) Hierarchical Agent Reasoning (H): A three-stage orchestration consisting of parallel diagnostic analysis (Stage 1), sequential dietary refinement (Stage 2), and safety-aware synthesis (Stage 3); (Right) Output Standardization (O): Projection of synthesized ADIME-structured insights into interoperable FHIR R4 clinical resources. We apply a chunking function휙:퐷 푔푢푖푑푒 →푐 1 , . . .,푐 푚 to segment documents into semantic units. Each chunk푐 푗 is mapped to an embedding vectore 푗 ∈R 푑 using the BGE-M3 model. For a given query푞, the retrieval function푟 푔푢푖푑푒 (푞)returns the top-푘 relevant chunks based on cosine similarity: 푟 푔푢푖푑푒 (푞)= arg max 퐶 ′ ⊂푐 푗 ,|퐶 ′ |=푘 ∑︁ 푐∈퐶 ′ e 푞 ·e 푐 ∥e 푞 ∥e 푐 ∥ (3) where푘=3 is selected to balance context window constraints with evidence sufficiency. 3.2.2 Structured Tool Retrieval. For medication analysis, we de- fine a discrete retrieval function푟 푑푟푢푔 (푀 푝 )that queries the Daily- Med/openFDA database. This function maps a drug set푀 푝 to a set of Structured Product Labels (SPLs), filtering for sections푆 푓표푑 ⊂ 푆푃퐿 related to drug-nutrient interactions. 3.3 Hierarchical Agent Reasoning (H) The inference process is modeled as a directed execution graph G 푒푥푒푐 =(A,퐸). The agent set is defined asA=푎 푚푔푟 ∪A 푑표푚푎푖푛 ∪ A 푟푒푓푖푛푒 ∪A 푠푦푛푡ℎ . Each agent푎 푖 is parameterized by a Large Lan- guage ModelM and a specific role prompt 휌 푖 . 3.3.1 Execution Topology & Rationale. We design the execution topology to mirror the clinical decision-making process for multi- morbidity management. We adopt a Parallel-then-Sequential archi- tecture (see Algorithm 1) grounded in three clinical principles: • Bias Mitigation via Independence (Parallel Analysis): In Stage 1, domain agents (A 푑표푚푎푖푛 ) process specific sub- spaces of the patient profile푋 푝 concurrently. This isolation ensures that the assessment of one condition (e.g., Diabetes) is not biased by the preliminary findings of another (e.g., Hypertension). This design simulates the independent assess- ment phase of specialized clinicians, effectively preventing the propagation of hallucinations or anchoring bias across domains. •Constraint Propagation (Sequential Refinement): In Stage 2, condition-specific dietitians (A 푟푒푓푖푛푒 ) refine the gen- eral dietary assessment based on the confirmed diagnoses from Stage 1. This models the clinical workflow of "Con- sultation & Adjustment," where a diagnosis acts as a condi- tional constraint on subsequent lifestyle interventions (e.g., a confirmed hypertension diagnosis activates DASH-specific sodium limits). •Multi-Objective Optimization (Synthesis): The final stage addresses the "multi-morbidity conflict" problem. By aggre- gating all upstream constraints, the system performs a syn- thesis that balances conflicting guidelines (e.g., renal vs. dia- betic diets), effectively solving a multi-objective optimization problem to maximize patient health utility under safety con- straints. NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA Algorithm 1 NutriOrion Inference Process Require: Patient Profile 푋 푝 , GuidelinesK Ensure: Personalized Plan 푌 푝 1: Initialize contextC ←∅ 2: Stage 1: Independent Assessment (Parallel) 3: for agent 푎 푖 ∈ A 푑표푚푎푖푛 do 4:Query 푞 푖 ← 푎 푖 .draft_query(푋 푝 ) 5:Evidence 퐸 푖 ←R(푞 푖 ,K) 6:Insight ℎ 푖 ← 푎 푖 .analyze(푋 푝 ,퐸 푖 ) 7: C ←C∪ℎ 푖 8: end for 9: Stage 2: Conditional Refinement (Sequential) 10: for specialist 푎 푠푝푒푐 ∈ A 푟푒푓푖푛푒 do 11:Context퐶 푠푝푒푐 ← SelectContext(C,푎 푠푝푒푐 ) 12: 푟 푠푝푒푐 ← 푎 푠푝푒푐 .reason(퐷 푝 ,퐶 푠푝푒푐 ,K 푠푝푒푐 ) 13: C ←C∪푟 푠푝푒푐 14: end for 15: Stage 3: Synthesis & Safety Verification 16: Priority Vector푉 푝푟푖표 ← 푎 푝푟푖표 (C) 17: Ranked Issues 퐿 ← Rank(푉 푝푟푖표 ) using Eq. (4) 18: Extract Hard DNI Constraints from Medication Agent 19: 퐶 푠푎푓푒 ← ExtractConstraints(C,푎 푚푒푑 ) 20: Generate plan strictly bounded by safety context 21: 푌 푝 ← 푎 푟푒푝표푟푡 (퐿,C,퐶 푠푎푓푒 ) 22: return 푌 푝 3.3.2 Algorithmic Flow. The formal inference procedure is detailed in Algorithm 1. Crucially, the process integrates both soft reason- ing (LLM-based) and hard safety checks via contextual constraint injection. The Safety Constraint Mechanism (Algorithm 1, lines 17-19) ensures pharmacological safety. Unlike post-hoc filtering which risks plan incoherence, this stage extracts contraindications (퐶 푠푎푓푒 ) from Stage 1 and injects them as hard negative constraints during the final synthesis. This guarantees that the generated plan푌 푝 is chemically valid by construction. 3.4 Multi-Objective Synthesis and Clinical Output (O) The synthesis stage aggregates the disparate insights from domain agents into a cohesive clinical plan. This process is governed by a Health Prioritization Agent and a Structured Projection Module. 3.4.1 Priority Scoring Mechanism. To resolve conflicting guidelines (e.g., DASH requiring high potassium vs. CKD requiring restriction), we define a scoring function푆(푗)for each identified health issue푗: 푆(푗)=푤 푠푒푣 · 휎 푠푒푣 (푗)+푤 푢푟푔 · 휎 푢푟푔 (푗)+푤 푚표푑 · 휎 푚표푑 (푗)(4) where휎 ∈ [0,1]represents normalized scores derived from clini- cal thresholds for Severity, Urgency, and Modifiability. Issues are ranked by 푆(푗) to determine the primary intervention focus. 3.4.2 ADIME-Structured Projection. Unlike generic LLMs that gen- erate free-text narratives, NutriOrion projects the synthesized in- sights into a strict ADIME (Assessment, Diagnosis, Intervention, Monitoring, Evaluation) schema. The intervention component of the plan,퐼 푝푙푎푛 ⊂ 푌 푝 , is structured as discrete, actionable directives: 퐼 푝푙푎푛 =(푎푐푡푖표푛, 푓표푑,푟푒푎푠표푛) | 푎푐푡푖표푛 ∈ Continue, Replace, Add (5) This structured classification—explicitly separates positive rein- forcement (Continue) from corrective substitutions (Replace)—aligns with the standardized terminology of the Nutrition Care Process (NCP). 3.4.3 Interoperability: FHIR-Aligned Output Projection. To bridge the gap between AI-driven generation and clinical deployment, we construct the output spaceYto be formally isomorphic to the HL7 FHIR R4NutritionOrderresource. We define a deter- ministic mapping functionΦ:푌 푝 → 푌 퐹퐻퐼푅 that transforms the finalized ADIME components into standardized, EHR-conformant artifacts. Specifically, agent-generated dietary protocols are pro- jected onto standardized clinical terminologies, while quantitative constraints—such as upper bounds on sodium intake—are encoded as structured nutrient definitions. A comprehensive specification is provided in Appendix A. By design, this architectural alignment renders NutriOrion inherently “FHIR-ready,” enabling seamless integration into hospital information systems. 3.5 Safety Verification Mechanism (G 푠푎푓푒 ) While hierarchical synthesis optimizes for clinical utility, patient safety requires strict adherence to pharmacological contraindica- tions. Instead of relying on probabilistic knowledge, we implement a Retrieval-Augmented Safety Constraint mechanism. Let퐶 푠푎푓푒 ⊂ Cbe the set of Drug-Nutrient Interactions (DNIs) identified by the Medication Agent using the DailyMed tool (Stage 1). During the final synthesis (Stage 3), these interactions are treated not as optional context, but as Hard Negative Constraints in the in- ference prompt. Formally, the synthesis function maximizes clinical utility푈(푌 푝 ) subject to: ∀푑 ∈ 퐶 푠푎푓푒 ,Sim(푌 푝 ,푑 푐표푛푡푟푎푖푛푑푖푐푎푡푖표푛 )< 휖(6) 4 Experiments We evaluate NutriOrion across five complementary dimensions: output structure and actionability (Section 4.2) measures ad- herence to standardized care formats and recommendation speci- ficity, medication safety (Section 4.3) ensures interventions avoid drug-nutrient interactions, clinical appropriateness (Section 4.4) assesses alignment with evidence-based guidelines, personaliza- tion (Section 4.5) quantifies biomarker-driven dietary tailoring, and food quality (Section 4.6) evaluates nutritional healthfulness and compositional improvements. These dimensions collectively validate clinical deployment readiness. 4.1 Experimental Setup 4.1.1 Dataset and Patient Cohort. We evaluated all 330 stroke sur- vivors from the NHANES 2011-2012 cycles [19]. Stroke survivors frequently present with multimorbidities, particularly hypertension and diabetes, making them ideal for evaluating personalized nu- trition intervention in complex clinical scenarios [48]. All patients had complete data required by NutriOrion, including dietary recall Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang (24-hour recall), laboratory biomarkers (e.g., HbA1c, blood pressure, lipid panel), and medication records. 4.1.2 Baselines and Ablation Studies. To rigorously evaluate Nu- triOrion(using GPT4o-mini), we benchmark against 12 baselines sharing identical tool access and output schemas. We categorize comparisons into three research questions (RQ). RQ1: Can Single Agents Replace Multi-Agent Architectures? We investigate whether a single powerful LLM can match multi- agent precision by evaluating three reasoning paradigms on open- weights models (Llama-3.1-70B, Qwen-2.5-72B): ReAct [51] (stan- dard dynamic thought-action loop), Chain-of-Thought (CoT) [45] (linear step-by-step reasoning without backtracking), and Self- Refine [25] (iterative Draft-Critique-Refine pipeline). Additionally, we benchmark against closed-source models (GPT-4.1, Claude-4- Sonnet) using ReAct to determine if foundation model capability overrides architectural benefits. RQ2: How Do Roles and Process Contribute? To disentangle NutriOrion’s core components, we conduct controlled ablations: w/o ADIME Structure retains 10 specialized agents but removes the DAG dependency to rely on dynamic manager orchestration; w/o Specialized Roles retains the ADIME workflow but replaces domain experts with generic "Clinical Nutritionist" agents to test the necessity of domain-specific backstories. RQ3: Is Structured Collaboration Superior? We compare against two prevalent multi-agent paradigms: Unstructured Collabora- tion (Round Table) simulating AutoGen-style [46] free discussion without a predefined pipeline, and Ensemble Voting simulating MedAgents-style [41] redundancy where parallel agents with di- verse personas (Conservative, Proactive, Evidence-Based) vote on recommendations. Table 1 summarizes all methods. Implementation details are provided in Appendix B. Table 1: Summary of baseline methods and ablations. All methods use identical tools and temperature = 0.0. MethodTypeLogic# AgentsCore Test NutriOrion (Ours)Multi-Agent Structured ADIME 10SOTA Performance + Specialized Roles ReAct [51]SingleDynamic Tool Loop1Dynamic Reasoning CoT [45]SingleLinear Planning1Linear Reasoning Self-Refine [25]SingleIterative Correction1Self-Correction API ReActSingleDynamic Tool Loop1Model Capability w/o ADIMEAblationDynamic Delegation10Workflow Impact w/o RolesAblationGeneric Roles + Fixed DAG10Persona Impact Round Table [46]BenchmarkUnstructured Discussion3Structure vs. Freedom Ensemble [41]BenchmarkParallel Persona Voting3+1Specialization vs. Redundancy 4.2 Output Structure & Actionability To ensure clinical validity beyond lexical metrics, we conducted expert evaluation using two Board-Certified Registered Dietitians (RDs). We employed the Nutrition Care Process Quality Evaluation and Standardization Tool (NCP-QUEST) [23], a validated instru- ment for assessing nutrition care documentation quality across the ADIME framework: Assessment, Diagnosis, Intervention, and Monitoring/Evaluation. Evaluators were blinded to model sources and independently scored each plan on NCP-QUEST’s 5-point Lik- ert scales, referencing evidence-based dietary guidelines including American Diabetes Association (ADA) standards and Dietary Ap- proaches to Stop Hypertension (DASH) protocols [2,11] to judge clinical appropriateness. Inter-rater agreement was substantial (Co- hen’s 휅= 0.82). Beyond structural completeness, we introduced a second qual- ity metric: Actionability Rate, which quantifies the proportion of recommendations containing specific, actionable food items versus generic dietary categories. For each plan, RDs annotated recommen- dations as either specific (e.g., "grilled salmon," "spinach") or generic (e.g., "healthy foods," "more protein"), and we calculated actionabil- ity as 1 minus the proportion of generic recommendations. This metric reveals whether models provide practically implementable guidance or merely abstract dietary principles Table 2 presents the dual-evaluation results. NutriOrion achieves an excellent expert rating (7.5/8) and state-of-the-art actionability (97.8%), effectively matching frontier commercial APIs like GPT-4.1 (91.2%) and significantly outperforming open-source baselines. Cru- cially, the actionability metric reveals a "fluency trap" in baseline models. For instance, while React-Qwen-72B achieves a respectable expert score (6.0/8) due to fluent reasoning, its actionability is crit- ically low (20.8%). This discrepancy is quantified by a moderate Pearson correlation (푟=0.63,푝<0.05) between metrics, confirm- ing that specificity is a necessary but insufficient condition for qual- ity. Models failing to produce structured ADIME outputs receive zero scores for intervention-specific dimensions despite sometimes generating reasonable diagnoses. NutriOrion’s specialized agent architecture successfully bridges the gap between abstract medical knowledge and actionable patient instructions, generating approxi- mately≈7 times less output than frontier models (2,663 vs. 17,925 average characters) while maintaining comparable clinical accuracy, making it more efficient for reviewing clinicians. Table 2: NCP-QUEST Scores & Actionability Rates. NCP- QUEST scores (0–8) averaged from two independent RDs (휅=0.82). Act.% (Actionability) denotes the success rate of generating concrete, identifiable food items (Calculated as 1− Non-specific Rate). ModelDiagnosis Interv. Logic Detail TotalAct.% Claude-Sonnet-42.02.01.51.57.091.2 GPT-4.12.01.51.51.56.591.2 NutriOrion2.02.02.01.57.597.8 React-Qwen-2.5-72B2.01.51.51.06.020.8 React-LLaMA-3-70B1.51.01.50.54.518.1 Models below produce non-structured outputs: Ablation-NoSpec.2.00.00.00.02.085.6 SelfRefine-LLaMA-3-70B1.50.00.00.01.543.4 SelfRefine-Qwen2.5-72B1.00.00.00.01.032.9 CoT-Qwen-2.5-72B1.00.00.00.01.07.8 CoT-LLaMA-3-70B1.00.00.00.01.0 26.4 Ablation-NoADIME0.00.00.00.00.024.2 Multiagent-RoundTable1.00.00.00.01.019.5 Multiagent-Ensemble0.00.00.00.00.037.6 4.3 Medication Safety Drug-food interactions represent a critical safety boundary in clini- cal nutrition, particularly for stroke patients managing polyphar- macy [8]. To assess this, we implemented a rigorous safety audit em- ploying a two-stage validation pipeline: (1) static screening against guideline-derived contraindication lists, followed by (2) ground- truth verification against the FDA DailyMed database [43], where NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA Table 3: Drug-Food Interaction Violations. Eval. (푁) denotes the number of outputs containing actionable food items. Rate is the percentage of outputs triggering FDA contraindications. Rows in gray indicate models with insufficient Actionability (<40% in Table 2). Their near-zero violation rates are artifacts of silence (failing to prescribe food) rather than safety. ModelEval. (푁)ViolationRate High Actionability Models (Act.> 85% in Table 2) Claude-Sonnet-41573211.5% GPT-4.11573516.6% Ablation-NoSpecialist1573416.6% NutriOrion (Ours)1322312.1% Moderate Actionability Baselines RoundTable114148.8% React-Qwen-2.5-72B12664.8% Ensemble15753.2% SelfRefine-Qwen-2.5155117.1% Low Actionability Artifacts (Act.< 40% in Table 2) Ablation-NoADIME15700.0% React-LLaMa-3-70B9700.0% CoT-Qwen-2.5-72B15710.6% CoT-LLaMA-3-70B15742.5% SelfRefine-LLaMA-315500.0% we queried specific drug labels to extract authoritative warning sections. We focused on three high-risk post-stroke cohorts: Warfarin users (푛=28), Potassium-sparing diuretic users (푛=15), and Statin users (푛=114). These groups were selected for their well- documented and potentially severe dietary interactions. The un- derlying pharmacological mechanisms, contraindicated food items, and corresponding guideline references are detailed in Appendix 5. Results. Table 3 reveals a systematic vulnerability we term the “Healthy Food Trap” : high-performing models (NutriOrion, Claude, GPT) frequently recommend generally healthy foods (e.g., leafy greens, bananas) that are specifically contraindicated for pharma- cotherapy subgroups, leading to violation rates of 11–17%. This occurs because nutritional heuristics (“eat vegetables”) conflict with context-specific drug constraints (e.g., Vitamin K restriction for Warfarin). While models like React-LLaMa-70B or Ablation-NoADIME ap- pear ostensibly safer with 0.0% violations, cross-referencing with Ta- ble 2 exposes this as an artifact of low Actionability (<40%). These models fail to prescribe concrete food items, outputting generic ad- vice (e.g., “eat healthy”) that bypasses safety filters by default. This passive safety is clinically useless as it fails to guide patient behav- ior. NutriOrion (12.1% violations) represents the only open-source architecture achieving active safety: it generates highly specific, actionable meal plans (Actionability 97.8% in Table 2) while main- taining a safety profile comparable to frontier closed-source models (Claude-Sonnet-4 at 11.5%). Failure Analysis. Detailed breakdowns in Appendix Table 5 reveal that Warfarin interactions account for the majority of violations (21–57% across high-actionability models). Despite NutriOrion’s deterministic Safety Gate, the residual 12% violation rate highlights a semantic gap in current pharmaceutical databases: explicit con- traindications in Product Labels (e.g., “avoid Vitamin K-rich foods”) do not always enumerate specific vegetables (spinach, kale), creat- ing a terminology mismatch that rule-based filters cannot resolve. This underscores the urgent need for comprehensive drug-nutrient Knowledge Graphs (KG) to augment LLM reasoning. 4.4 Clinical Appropriateness To assess holistic clinical utility, the same two Registered Dieti- tians (RDs) from Section 4.2 evaluated 30 randomly selected Nu- triOrion interventions using a structured survey instrument (Ap- pendix C). The survey aligns with NCP domains and post-stroke guidelines [21], comprising six dimensions: (A) Guideline Align- ment, (B) Dietary Components, (C) Balance, (D) Clarity, (E) Safety (identification of unsafe foods), and (F) Overall Quality. Results. NutriOrion demonstrated strong clinical adherence, scoring highest in Guideline Alignment (Mean=4.69/5) and Dietary Components (4.66). Dietary Balance (4.51) and Safety (4.28) also achieved high ratings, confirming that the system effectively ad- dresses multimorbidity requirements. Figure 5 in the appendix visualizes the evaluation profile. Notably, Overall Quality scored 3.87. Open-text feedback clari- fied that while food recommendations were chemically accurate (Section 4.5) and specific (Section 4.2), the lack of "practical meal preparation guidance (e.g., recipes, culinary preparation methods)" lowered the adoption score. This distinguishes a logistical gap from the specificity gap solved by our architecture, highlighting that future clinical LLMs must extend beyond "what to eat" to "how to prepare and quantify it." Despite this gap, quantitative analysis confirms clinical viability: 81.4% of plans achieved clinical acceptability, and 64.4% of RDs indicated willingness to sign off on them (Adoption Intention). Re- garding safety, while 85.6% of items received positive ratings, 37.3% of evaluations noted missed subtle unsafe foods, corroborating the "Healthy Food Trap" findings in Section 4.3. Inter-annotator agreement varied substantially: highest for cat- egorization tasks (79–97%) but lowest for absolute safety red flag counts (8%). Qualitative audit reveals this stems from variance in an- notation granularity rather than clinical dissent: one RD tended to flag individual contraindicated ingredients (e.g., "salt", "soy sauce" →count=2), while the other flagged the composite meal (e.g., "high sodium dish"→count=1). Crucially, evaluators achieved high con- sensus on the presence of safety risks, validating the overall safety scores despite the counting ambiguity. 4.5 Responsiveness to Patient Profiles To quantify personalization in generated nutrition interventions, we measured the alignment between patient health biomarkers and the nutritional composition of recommended foods. We mapped food items in generated interventions to Food and Nutrient Database for Dietary Studies (FNDDS) codes, the same coding system NHANES uses to obtain nutritional data, using the method from [47]. This enabled extraction of complete nutritional profiles for each rec- ommended food. We selected five key health biomarkers: systolic blood pressure (BP), glucose, HbA1c, total cholesterol, and BMI. For each patient, we calculated Pearson correlation coefficients between health biomarkers and corresponding dietary nutrients: Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang (1) systolic BP vs. sodium content, (2) glucose vs. sugar content, (3) HbA1c vs. sugar content, (4) total cholesterol vs. fat content, and (5) BMI vs. energy content. The personalization hypothesis is that appropriate recommen- dations should exhibit negative correlations: when patients have elevated biomarkers (e.g., high BP), recommended foods should contain lower levels of risk nutrients (e.g., sodium). As visualized in Figure 3, negative values (Blue) indicate high personalization quality, while positive values (Red) imply safety failures. NutriOr- ion demonstrates superior personalization by consistently achiev- ing the strongest negative correlations across all five metrics (e.g., HbA1c:−0.35; Cholesterol:−0.31), appearing as the distinct deep blue region in the heatmap. In contrast, baseline models frequently exhibit unsafe positive correlations (red zones). Complete correla- tion results and comparative analysis are presented in Appendix E. 4.6 Food Quality Assessment To evaluate the nutritional quality of recommended foods, we em- ployed Food Compass 2.0 [6], a comprehensive food healthfulness ranking system that scores foods based on multiple nutritional attributes including nutrient composition, food processing level, and health outcomes associated with consumption. Food Compass 2.0 provides a rating score ranging from 0 to 100 for each FNDDS food code, with higher scores indicating greater nutritional quality. We leveraged the FNDDS food code mapping from Section 4.5 to obtain Food Compass scores for each food item in the generated nutrition interventions. For each intervention, we calculated the average Food Compass score across all recommended food items, then aggregated these scores across all patient samples to obtain a SBPGlucoseHbA1cChol.BMI Health Metrics Ablation-NoSpec. SelfRefine-LLaMA-3-70B CoT-LLaMA-3-70B Ablation-NoADIME Multiagent-RoundTable Multiagent-Ensemble Claude-Sonnet-4 GPT-4.1 NutriOrion React-Qwen-2.5-72B React-LLaMA-3-70B 0.770.58-0.21-1.000.42 -0.27-0.84 0.060.810.360.500.10 0.620.32-0.06-0.27-0.09 0.65-0.24 0.410.900.780.01-0.34 -0.11-0.050.040.17-0.03 0.10-0.14-0.140.050.12 -0.26-0.30-0.35-0.31-0.19 -0.08-0.25 -0.030.02 −1.00 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 Pearson Correlation (Lower/Blue is Better) Figure 3: Personalization Heatmap. Pearson correlations be- tween patient biomarkers and nutrient intake. Blue (Neg- ative) indicates effective personalization (e.g., High BP→ Low Sodium), while Red (Positive) implies unsafe recom- mendations. NutriOrion consistently achieves the strongest protective correlations across all five metrics, significantly outperforming baselines. 0.00.20.40.60.81.0 Actionability (Success Rate) 40 50 60 70 80 90 Food Compass Score 2.0 (FCS2) GPT-4.1 Claude-Sonnet-4 React-LLaMA-3-70B React-Qwen2.5-72B CoT-LLaMA-70B CoT-Qwen-72B SelfRefine-LLaMA-70B SelfRefine-Qwen-72B Multiagent-Ensemble Multiagent-RoundTable Ablation-NoADIME Ablation-NoSpec. NutriOrion 2.2 2.4 2.6 2.8 3.0 3.2 3.4 NutriScore (Color) Figure 4: Multi-dimensional Performance (Bubble Chart). X-axis: Actionability; Y-axis: Food Compass Score 2.0 (Nu- tritional Quality); Bubble Size: Dietary Diversity; Color: Nu- triScore Health Gradient (Red→Green). NutriOrion (top- right) achieves the optimal balance: high quality, actionable output, and high diversity. Gray labels indicate unstructured baselines, as detailed in Table 2. mean nutritional quality score for each model. Figure 4 presents a multi-dimensional performance comparison, plotting actionabil- ity (x-axis) against Food Compass Score 2.0 (y-axis), with bubble size representing dietary diversity (unique food codes) and color indicating NutriScore-based health quality (green denotes higher scores). Figure 4 visualizes the performance landscape, revealing distinct architectural behaviors. NutriOrion and GPT-4.1 occupy the Opti- mal Frontier (Top-Right), achieving both high quality (FCS≈80) and high actionability (>0.85), proving that nutritional density can co- exist with specific planning. In contrast, baselines like React-LLaMA and RoundTable fall into a "Conservative Trap" (Top-Left), achieving high quality scores (≈75–78) but suffering from low Actionability (<0.3). This indicates a form of "Passive Quality", they achieve high scores by recommending generic healthy categories rather than specific meals, limiting their clinical utility. Meanwhile, CoT-based models mostly reside in the Failure Zone (Bottom-Left), failing in both dimensions with low-quality outputs (FCS<50, Red Bubble) and negligible actionability. The bubble size reveals that NutriOrion recommends a diverse range of foods, indicating varied recommen- dations tailored to individual patient contexts rather than repetitive suggestions. The combination of high food quality, high actionabil- ity, substantial diversity, and strong NutriScore demonstrates that NutriOrion successfully balances multiple competing objectives in nutrition intervention generation. NutriOrion’s recommendations demonstrate clinically meaningful dietary improvements: reducing energy by 8%, sodium by 9%, and sugars by 12%, while simultane- ously increasing beneficial nutrients with potassium up 27% and NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA fiber up 167%. These targeted modifications align with dietary guide- lines for chronic disease management. Comprehensive nutritional composition analysis is presented in Appendix F. 5 Conclusion Our evaluation of NutriOrion on 330 multimorbid stroke patients demonstrates that specialized multi-agent architectures fundamen- tally outperform monolithic reasoning across safety, personaliza- tion, and clinical appropriateness dimensions, with controlled abla- tions revealing that both role specialization and structured ADIME workflows are necessary, as removing either degrades performance to baseline levels. The framework achieves 97.8% actionability in generating specific food recommendations while maintaining 12.1% drug-food violation rates comparable to frontier commercial APIs, with registered dietitians rating 81.4% of plans as clinically accept- able and demonstrating robust personalization through consistent negative correlations (ranging from -0.26 to -0.35) between patient biomarkers and corresponding dietary risk nutrients across all eval- uated health metrics. Building upon these foundational results, future work will extend the system in two important directions: incorporating precise portion size specifications by integrating standardized serving databases to provide quantitative nutrient in- take control, and conducting prospective clinical trials to evaluate patient-reported outcomes including adherence rates, satisfaction scores, and longitudinal health improvements in real-world deploy- ment settings. More broadly, NutriOrion enables on-demand dietary support accessible anywhere, increasing the reach of personalized nutrition intervention and allowing populations worldwide to ben- efit from advanced LLM capabilities and digital health tools for managing chronic diseases. Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang References [1]Giuseppe Agapito, Mariadelina Simeoni, Barbara Calabrese, Ilaria Caré, Theodora Lamprinoudi, Pietro H Guzzi, Arturo Pujia, Giorgio Fuiano, and Mario Cannataro. 2018. DIETOS: A dietary recommender system for chronic diseases monitoring and management. Computer methods and programs in biomedicine 153 (2018), 93–104. [2] American Diabetes Association Professional Practice Committee. 2024. Intro- duction and Methodology: Standards of Care in Diabetes—2024. Diabetes Care 47, Supplement_1 (Jan. 2024), S1–S4. doi:10.2337/dc24-SINT Abbreviated; full Standards in same supplement. [3]Maryam Amiri, Fatemeh Sarani Rad, and Juan Li. 2024. Delighting Palates with AI: Reinforcement Learning’s Triumph in Crafting Personalized Meal Plans with High User Acceptance. Nutrients 16, 3 (2024), 346. [4]Nishanthi Anandabaskar. 2019. Drug information. In Introduction to Basics of Pharmacology and Toxicology: Volume 1: General and Molecular Pharmacology: Principles of Drug Action. Springer, 223–234. [5] Elham Asgari, Nina Montaña-Brown, Magda Dubois, Saleh Khalil, Jasmine Bal- loch, Joshua Au Yeung, and Dominic Pimenta. 2025. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine 8, 1 (2025), 274. [6]Eden M Barrett, Peilin Shi, Jeffrey B Blumberg, Meghan O’Hearn, Renata Micha, and Dariush Mozaffarian. 2024. Food Compass 2.0 is an improved nutrient profiling system to characterize healthfulness of foods and beverages. Nature Food 5, 11 (2024), 911–915. [7] Andrew A Borkowski and Alon Ben-Ari. 2025. Multiagent AI Systems in Health Care: Envisioning Next-Generation Intelligence. Federal Practitioner 42, 5 (2025). [8]Rabia Bushra, Nousheen Aslam, and Arshad Yar Khan. 2011. Food-drug interac- tions. Oman Medical Journal 26, 2 (2011), 77–83. [9]Jiaqi Chen, Yu Fang, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al.2024. MedRAG: Medical Retrieval- Augmented Generation for Better Health Question Answering. arXiv preprint arXiv:2402.09928 (2024). [10]Saifur Rahman Chowdhury, Dipak Chandra Das, Tachlima Chowdhury Sunna, Joseph Beyene, and Ahmed Hossain. 2023. Global and regional prevalence of multimorbidity in the adult population in community settings: a systematic review and meta-analysis. EClinicalMedicine 57 (2023). [11]DASH Collaborative Research Group. 1997. A clinical trial of the effects of dietary patterns on blood pressure. New England Journal of Medicine 336, 16 (April 1997), 1117–1124. doi:10.1056/NEJM199704173361601 [12]Shimaa Elkomy and Tim Jackson. 2024. WHO non-communicable diseases Global Monitoring Framework: Pandemic resilience in sub-Saharan Africa and Low-income Countries. Socio-Economic Planning Sciences 95 (2024), 102043. doi:10.1016/j.seps.2024.102043 [13]David Elsweiler, Christoph Trattner, and Morgan Harvey. 2017. Exploiting food choice biases for healthier recipe recommendation. In Proceedings of the 40th international acm sigir conference on research and development in information retrieval. 575–584. [14] Ramón Estruch, Emilio Ros, Jordi Salas-Salvadó, Maria-Isabel Covas, Dolores Corella, Fernando Arós, Enrique Gómez-Gracia, Valentina Ruiz-Gutiérrez, Miquel Fiol, José Lapetra, et al.2018. Primary prevention of cardiovascular disease with a Mediterranean diet supplemented with extra-virgin olive oil or nuts. New England journal of medicine 378, 25 (2018), e34. [15]Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023). [16]Rozenn Gazan, Chloé MC Brouzes, Florent Vieux, Matthieu Maillot, Anne Lluch, and Nicole Darmon. 2018. Mathematical optimization to explore tomorrow’s sustainable diets: a narrative review. Advances in Nutrition 9, 5 (2018), 602–616. [17]Diabetes Prevention Program Research Group. 2002. Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. New England journal of medicine 346, 6 (2002), 393–403. [18]International Confederation of Dietetic Associations. 2021. Education and Work Report 2021: Dietitian-Nutritionists Around the World. ICDA report. [19]Clifford Leroy Johnson, Sylvia M Dohrmann, Vicki L Burt, and Leyla Kheradmand Mohadjer. 2014. National health and nutrition examination survey: sample design, 2011-2014. Number 2014. US Department of Health and Human Services, Centers for Disease Control and . . . . [20]Kyriakos Kalpakoglou, Lorena Calderón-Pérez, Noemi Boqué, Metin Guldas, Çağla Erdoğan Demir, Lazaros P Gymnopoulos, and Kosmas Dimitropoulos. 2025. An AI-based nutrition recommendation system: technical validation with insights from Mediterranean cuisine. Frontiers in Nutrition 12 (2025), 1546107. [21]Dawn O. Kleindorfer, Amytis Towfighi, Seemant Chaturvedi, Kevin M. Cockroft, Jose Gutierrez, Dawn Lombardi-Hill, Hooman Kamel, Walter N. Kernan, Steven J. Kittner, Thabele M. Leslie Nguyen, and Linda S. Williams. 2021. 2021 Guideline for the Prevention of Stroke in Patients With Stroke and Transient Ischemic Attack: A Guideline From the AHA/ASA. Stroke 52, 7 (July 2021), e364–e467. doi:10.1161/STR.0000000000000375 [22]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al.2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 (2020), 9459–9474. [23]Sherri L Lewis, Leslie S Miranda, Julie Kurtz, Linda M Larison, W James Brewer, and Constantina Papoutsakis. 2022. Nutrition Care Process Quality Evaluation and Standardization Tool: The Next Frontier in Quality Evaluation of Documen- tation. Journal of the Academy of Nutrition and Dietetics 122, 3 (2022), 650–660. doi:10.1016/j.jand.2021.07.004 [24]Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the association for computational linguistics 12 (2024), 157–173. [25]Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2023. Self-Refine: Iterative Refinement with Self-Feedback. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 36. [26] Ahmed A Metwally, Ariel K Leong, Aman Desai, Anvith Nagarjuna, Dalia Perel- man, and Michael Snyder. 2021. Learning personal food preferences via food logs embedding. In 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2281–2286. [27] Weiqing Min, Shuqiang Jiang, Linhu Liu, Yong Rui, and Ramesh Jain. 2019. A survey on food computing. Acm Computing Surveys (CSUR) 52, 5 (2019), 1–36. [28]Marina Padovan, Fernando Ribeiro de Senna, Juliana Klein Kimura, Samara Tor- torella Nascimento, Antonio Carlos Moretti, and Caroline Dário Capitani. 2023. Optimized menu formulation to enhance nutritional goals: design of a mixed inte- ger programming model for the workers’ food program in Brazil. BMC nutrition 9, 1 (2023), 51. [29]Ilias Papastratis, Dimitrios Konstantinidis, Petros Daras, and Kosmas Dimitropou- los. 2024. AI nutrition recommendation using a deep generative model and ChatGPT. Scientific Reports 14, 1 (2024), 14620. [30] Risha I Patel and Robert D Beckett. 2016. Evaluation of resources for analyzing drug interactions. Journal of the Medical Library Association: JMLA 104, 4 (2016), 290. [31] Pengcheng Qiu, Chaoyi Wu, Shuyu Liu, Yanjie Fan, Weike Zhao, Zhuoxia Chen, Hongfei Gu, Chuanjin Peng, Ya Zhang, Yanfeng Wang, et al.2025. Quantifying the reasoning abilities of LLMs on clinical cases. Nature Communications 16, 1 (2025), 9799. [32] David Ribeiro, João Machado, Jorge Ribeiro, Maria João M Vasconcelos, Elsa F Vieira, and Ana Correia De Barros. 2017. SousChef: mobile meal recommender system for older adults. In International Conference on Information and Commu- nication Technologies for Ageing Well and e-Health, Vol. 2. SCITEPRESS, 36–45. [33]Mehrdad Rostami, Usman Muhammad, Saman Forouzandeh, Kamal Berahmand, Vahid Farrahi, and Mourad Oussalah. 2022. An effective explainable food rec- ommendation using deep image clustering and community detection. Intelligent Systems with Applications 16 (2022), 200157. [34]Dimitri Roustan, François Bastardot, et al.2025. The clinicians’ guide to large language models: A general perspective with a focus on hallucinations. Interactive journal of medical research 14, 1 (2025), e59823. [35]Nicholas K Schiltz. 2022. Prevalence of multimorbidity combinations and their association with medical costs and poor health: a population-based study of US adults. Frontiers in Public Health 10 (2022), 953886. [36]Sina Shool, Sara Adimi, Reza Saboori Amleshi, Ehsan Bitaraf, Reza Golpira, and Mahmood Tara. 2025. A systematic review of large language model (LLM) evaluations in clinical medicine. BMC Medical Informatics and Decision Making 25, 1 (2025), 117. [37]Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. 2023. Large language models encode clinical knowledge. Nature 620, 7972 (2023), 172–180. [38]George Siopis, Alexandra Jones, and Margaret Allman-Farinelli. 2020. The dietetic workforce distribution geographic atlas provides insight into the inequitable access for dietetic services for people with type 2 diabetes in Australia. Nutrition & Dietetics 77, 1 (2020), 121–130. [39]Annalynn Skipper. 2007. Applying the nutrition care process: nutrition diagnosis and intervention. Support Line 29, 6 (2007), 12–23. [40]Xiangru Tang, Yilun An, Zexin Jiang, Yuwei Wu, Yuxin Wang, Bill Yuchen Lin, Zhiming Lu, Yujie Wang, Jinyi Wang, Hanyu Zhu, et al.2023. MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning. arXiv preprint arXiv:2311.10537 (2023). [41]Xiangru Tang, Anni Zou, Zhuosheng Zhang, Yilun Zhao, Xingang Zhang, Arman Cohan, and Mark Gerstein. 2023. MedAgents: Large Language Models as Clinical Reasoners. arXiv preprint arXiv:2311.10537 (2023). [42]Thi Ngoc Trang Tran, Müslüm Atas, Alexander Felfernig, and Martin Stettinger. 2018. An overview of recommender systems in the healthy food domain. Journal of Intelligent Information Systems 50, 3 (2018), 501–526. [43]U.S. National Library of Medicine. 2024. DailyMed. https://dailymed.nlm.nih.gov/ NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA [44]Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al.2024. A Survey on Large Language Model based Autonomous Agents. Frontiers of Computer Science (2024), 1–26. [45]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 35. 24824–24837. [46] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Wang, and Shaokun Zhang. 2023. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv preprint arXiv:2308.08155 (2023). [47]Runze Yan, Hanqi Luo, Jiaying Lu, Darren Liu, Hannah Posluszny, Mehak Preet Dhaliwal, Janice MacLeod, Yao Qin, Carl Yang, Terry J Hartman, et al.2025. Di- etAI24 as a framework for comprehensive nutrition estimation using multimodal large language models. Communications Medicine 5, 1 (2025), 458. [48]Shaoyi Yan, Shuo Sha, Senjie Li, Di Wang, and Yongping Jia. 2023. Association between hypertension and stroke in US adults in the National Health and Nutri- tion Examination Surveys (NHANES) 2007 to 2018. Postgraduate Medicine 135, 2 (2023), 187–194. [49]Eric Yang, Tomas Garcia, Hannah G Williams, Bhawesh Kumar, Martin Ramé, Eileen Rivera, Yiran Ma, Jonathan Amar, Caricia Catalani, and Yugang Jia. 2025. A Behavioral Science-Informed Agentic Workflow for Personalized Nutrition Coaching: Development and Validation Study. JMIR Formative Research 9 (2025), e75421. [50]Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Anthony B Costa, Mona G Flores, et al.2023. Large language models for automated clinical note generation. npj Digital Medicine 6, 1 (2023), 231. [51]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations (ICLR). https: //openreview.net/forum?id=WE_vluYUL-X [52]Cyril Zakka, Rushabh Shad, Akash Chaurasia, Alex R Dalal, Jennifer L Kim, Michael Moor, Kevin Chang, Kevin Fong, Jeffrey Boyd, Pujan Gupta, et al.2024. Almanac: Retrieval-augmented language models for clinical medicine. NEJM AI 1, 2 (2024), AIoa2300068. Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang A FHIR Interoperability Schema To demonstrate the clinical deployment readiness of NutriOrion, we provide the explicit mapping schema between our system’s internal ADIME-structured JSON output and the HL7 FHIR R4 NutritionOrderresource standard. This mapping ensures that the AI-generated interventions can be directly ingested by Electronic Health Record (EHR) systems without loss of semantic fidelity. Table 4: Full Mapping Schema: From NutriOrion ADIME Out- put to FHIR R4 NutritionOrder NutriOrion Internal Field ADIME Stage FHIR R4 Target Element hypertension_diet (e.g., DASH) AssessmentoralDiet.type.coding (e.g., SNOMED: 182922004) diabetes_diet (e.g., CarbCount) AssessmentoralDiet.type.coding nutrient_flags.sodiumDiagnosis oralDiet.nutrient.modifier (e.g., LOINC) nutrient_flags.limit_val Diagnosis oralDiet.nutrient.amount intervention.replace InterventionoralDiet.exclude- FoodModifier intervention.addInterventionoralDiet.food- PreferenceModifier medication.supplementInterventionsupplement.type health_priority.rankDiagnosis priority (Routine/Urgent) monitoring.frequencyMonitoringoralDiet.schedule The mapping functionΦis implemented as a deterministic parser that validates the agent’s output against the FHIR structure defini- tion before final delivery. B Implementation Details of Baselines Experiments were conducted using vLLM for open-source models and official APIs for closed-source models. All agents operate at temperature 0.0 to ensure reproducibility in clinical reasoning. B.1 Single-Agent Configurations The single agent acts as a "Comprehensive Specialist". • ReAct: Standard loop withmax_iter=50. We implemented aNoOpToolto prevent early termination failures common in open-source models. • CoT: Enforces a strict linear sequence: (1) Body Metrics→ (2) Guidelines→(3) Meds→(4) Diet→(5) Report. No tool looping allowed. •Self-Refine: A 3-stage chain sharing the LLM backbone: (1) Draft (tool-enabled), (2) Critique (text-only,<200 words to prevent context overflow), and (3) Refine (tool-enabled). B.2 Multi-Agent Baseline Configurations All multi-agent baselines use GPT-4o-mini as the backbone. •Ablations: w/o ADIME uses a hierarchical manager with the prompt: "There is no predefined workflow. Coordinate as you see fit." w/o Roles homogenizes all agent prompts to a generic "Clinical Nutritionist" while keeping the DAG. • Round Table (AutoGen-style): Consists of Physician, Di- etitian, and Pharmacist withallow_delegation=Trueto simulate free-form MDT meetings. A strict timeout is en- forced to prevent infinite loops. • Ensemble (MedAgents-style): Three parallel agents with distinct personas (Conservative, Proactive, Evidence-Based) analyze independently. An Aggregator Agent merges out- puts using majority voting for recommendations and union logic for safety warnings. C Registered Dietitian Evaluation Survey Fig. 5RD Clinical Appropriateness Evaluation. Mean scores (1–5 scale) across six dimensions for NutriOrion out- puts Below is the full content of the survey used for the dietitian review. Evaluation Survey Instructions. Please evaluate the given nutrition plan for this spe- cific patient using your professional judgment. Use a 5-point Likert scale unless otherwise specified: 1 = Strongly Disagree, 2 = Disagree, 3 = Neutral, 4 = Agree, 5 = Strongly Agree. Provide N/A if a statement is not applicable, or insufficient information is available. Section A: Patient Profile & Guideline Alignment (GA) GA1.The plan is consistent with the patient’s health profile, including diagnoses, anthropometrics, medications, aller- gies/intolerances. (1–5 / N/A) GA2. Overall, the plan aligns with applicable evidence-based guidelines for post-stroke care (e.g., ADA, DASH). (1–5) Section B: Appropriateness for Key Dietary Components (AP) Stem: “For the following dietary components, the AI’s restric- tion/intake recommendation is appropriate given the patient’s condition and guidelines.” Rate each item: AP1. Sodium (1–5 / N/A) AP2.Total Fat (considering saturated/trans-fat and cholesterol as part of total fat) (1–5 / N/A) AP3. Total Energy (calories) (1–5 / N/A) AP4. Added Sugars (1–5 / N/A) AP5. Dietary Fiber (1–5 / N/A) Section C: Dietary Balance (DB) DB1.The overall dietary pattern is balanced across food groups (e.g., fruits/vegetables, whole grains, low-fat dairy, lean pro- teins, healthy fats). (1–5) Section D: Clarity & Actionability (CL) CL1. The patient-facing summary is clear, concise, and under- standable. (1–5) NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines Conference’17, July 2017, Washington, DC, USA CL2.The “Why” explanations are accurate and evidence-informed. (1–5) CL3. The diet orders/instructions (limits/substitutions/additions) are specific and actionable (e.g., portions, frequency, concrete examples). (1–5) Section E: Safety (SF) Answer Yes/No. If “No,” a brief comment is required. Safety Red Flags (Definition):Any recommendation that con- flicts with the patient’s diagnoses, medications, or aller- gies/intolerances; any intake that exceeds/under-shoots guide- line thresholds posing clear risk; or any explicit drug–nutrient– disease incompatibility. SF1.Identification: The plan correctly identifies unsafe foods in the patient’s existing diet. (Yes/No) If No, please explain: SF2.Substitutions: Proposed substitute foods are functionally comparable (meet the same dietary intent) and safe. (Yes/No) If No, please explain: SF3.Additions: Newly recommended foods are safe and effective for the patient. (Yes/No) If No, please explain: SF4.Contraindications: The plan does not introduce drug– nutrient or disease-related contraindications. (Yes/No) If No, please explain: SF5.Number of Safety Red Flags identified:(integer; 0 if none) Section F: Overall Quality (OQ) OQ1. Overall, this AI-generated plan is clinically acceptable for this patient. (1–5) OQ2.(Adoption intention) I would be comfortable signing/co- signing this plan. (1–5) OQ3. Other comments or risks not covered above: Section G: Workload (WL) WL1. Time spent reviewing this plan (minutes): Section H: Open Comments D NCP-QUEST Evaluation Criteria In our expert evaluation (Section 4.2), Registered Dietitians assessed the generated nutrition documentation based on the following four dimensions derived from the NCP-QUEST framework [23]: •Nutrition Diagnosis (D): Evaluates the accuracy in identifying patient-specific nutritional problems. A score of 2 requires the explicit prioritization of the most urgent nutritional issue based on the provided clinical data. •Nutrition Intervention (I): Assesses the clinical appropriate- ness of the structured actions. Recommendations must use stan- dard terminology (e.g., “continue,” “replace,” “add”) to receive a full score. • D→I Link: Measures the logical consistency between the di- agnosis and the prescribed intervention. The intervention must Dietary Balance (4.51) Dietary Components Appropriateness (4.66) Guideline Alignment (4.69) Overall Quality (3.87) Safety (4.28) Clarity & Actionability (4.24) 1 2 3 4 5 Figure 5: RD Clinical Appropriateness Evaluation. Mean scores (1–5 scale) across six dimensions for NutriOrion out- puts directly address the etiology or signs/symptoms identified in the diagnosis. •Specificity: Evaluates the level of actionable detail. A score of 2 is awarded only if recommendations prescribe specific food items (e.g., “grilled salmon,” “spinach”) rather than generic categories (e.g., “healthy protein,” “vegetables”). Each dimension is scored on a discrete scale: •0 (Not Documented): The criteria are missing or incorrect. •1 (Partially Documented): The criteria are present but lack specific detail or full clinical alignment. • 2 (Adequately Documented): The criteria are fully met with specific, actionable, and clinically accurate details. E Personalization Correlation Analysis Figure 3 presents a heatmap visualization of Pearson correlations between patient health biomarkers (systolic BP, glucose, HbA1c, total cholesterol, BMI) and the corresponding dietary nutrients in recommended foods (sodium, sugar, sugar, fat, calories respec- tively) across different models. The color scale ranges from dark blue (strong negative correlation) to dark red (positive correlation), where more negative values indicate better personalization, as a well-personalized system should reduce risk nutrients when pa- tients have elevated health biomarkers. NutriOrion (highlighted) consistently achieves the strongest negative correlations across four of five health metrics compared to all baselines, with cor- relation coefficients ranging from -0.26 to -0.35 for systolic BP, glucose, HbA1c, and total cholesterol, indicating robust personal- ization where the multi-agent framework successfully interprets these health biomarkers and adapts dietary recommendations ac- cordingly. Notably, several baseline models exhibit positive corre- lations for certain metrics (e.g., CoT-LLaMA-3-70B shows 0.81 for glucose, Multiagent-Ensemble shows 0.78 for HbA1c), which repre- sents unsafe recommendations that could worsen health conditions Conference’17, July 2017, Washington, DC, USAJunwei Wu, Runze Yan, Hanqi Luo, Darren Liu, Minxiao Wang, Kimberly L. Townsend, Lydia S. Hartwig, Derek Milketinas, Xiao Hu, and Carl Yang Table 5: Drug-Food Interaction Violations by Clinical Scenario (DailyMed-based). The three scenarios represent high-risk drug-food interactions prevalent in stroke patients. Warfarin users face the highest violation rates across all high-quality models due to the “healthy food trap” where vitamin K-rich vegetables (spinach, broccoli, kale) are commonly recommended. Warfarin (n=28)K-Sparing Diuretics (n=15)Statins + Stroke (n=114) ModelEvalViolRateEvalViolRateEvalViolRate Claude-Sonnet-428621.4%151173.3%11410.9% NutriOrion26934.6%12758.3%9400.0% GPT-4.1281346.4%151173.3%11421.8% Ablation-NoSpec. 281657.1%151066.7%11400.0% Single-Qwen-72B 23521.7%10110.0%9300.0% SelfRefine-Qwen 28725.0%15320.0%11210.9% RoundTable21628.6%12433.3%8100.0% Note: The potassium-sparing diuretic scenario shows uniformly high violation rates (20–73%) because bananas, avocados, and tomatoes—common “heart-healthy” recommendations—are contraindicated for hyperkalemia risk. by suggesting foods high in risk nutrients for patients with elevated biomarkers. For BMI, NutriOrion shows -0.19 correlation, which is not the strongest among compared methods, as SelfRefine-LLaMA- 3-70B (-0.84) and Multiagent-Ensemble (-0.34) demonstrate stronger negative correlations; however, as discussed in Section 4.2, these models produced substantially lower actionability rates and gen- erated fewer specific food recommendations, resulting in limited samples for correlation calculation, making their BMI correlations potentially unrepresentative of consistent personalization capa- bility. Overall, NutriOrion’s balanced performance across all five health metrics demonstrates its ability to provide comprehensive, safe personalization for multimorbid patients. Energy (kcal)Sodium (mg) Sugars (g) Potassium (mg)Fiber (g) 0 50 100 150 200 250 Amount 195.27 219.38 10.6 214.1 1.67 179.75 200.21 9.38 271.74 4.46 Original IntakeNutriOrion RecommendedPositive ChangeNegative Change 0 25 50 75 100 125 150 175 Change (%) -8% -9% -12% +27% +167% Figure 6: Average dietary nutrients from original foods and NutriOrion recommendations. Bars show absolute amounts, and markers indicate percent change relative to original in- take. F Nutritional Composition Changes Figure 6 compares the average dietary nutrient content between patients’ original food intake (blue bars) and NutriOrion’s recom- mended foods (green bars) across five key nutrients, with bars show- ing absolute amounts and markers indicating percent change rela- tive to original intake (green upward triangles for positive changes, red downward triangles for negative changes). NutriOrion’s rec- ommendations demonstrate clinically meaningful dietary modifi- cations aligned with chronic disease management guidelines by reducing intake of risk-associated nutrients: energy decreases by 8% (from 195.27 to 179.75 kcal), sodium decreases by 9% (from 219.38 to 200.21 mg), and sugars decrease by 12% (from 10.6 to 9.38 g), which are critical reductions for managing obesity, hypertension, and diabetes respectively. Simultaneously, NutriOrion substantially increases intake of beneficial nutrients, with potassium increasing by 27% (from 214.1 to 271.74 mg) and dietary fiber increasing by 167% (from 1.67 to 4.46 g); higher potassium intake benefits both hypertension and cardiovascular health by counteracting sodium’s effects, while increased fiber improves glycemic control in diabetes and promotes satiety for weight management. These nutrient shifts demonstrate that NutriOrion does not simply restrict intake across all categories but rather makes targeted, evidence-based adjust- ments, reducing risk-associated nutrients while simultaneously increasing protective nutrients, reflecting sophisticated understand- ing of dietary guidelines for multimorbidity management. The sub- stantial fiber increase (167%) is particularly noteworthy, as fiber is often deficient in typical diets and plays multiple beneficial roles in chronic disease management, validating that NutriOrion generates actionable dietary interventions that translate high-level clinical objectives into concrete, beneficial nutritional changes.