HealthGenie: Empowering Users with Healthy Dietary Guidance through Knowledge Graph and Large Language Models

TL;DR

HealthGenie addresses the challenge of delivering personalized dietary guidance by coupling a nutrition-focused knowledge graph with large language models in a circular, interactive workflow. The system grounds LLM outputs in a curated KG, visualizes relationships for quick overview, and enables users to iteratively refine recommendations through inclusion/exclusion and query generation. A formative study informs design goals (visualization, explainability, and adaptive personalization), followed by a within-subject user study (N=12) showing HealthGenie improves information organization, task efficiency, and perceived usefulness compared with a baseline, while highlighting challenges in data coverage, latency, and KG structure. The work demonstrates the potential of LLM-KG integration for explainable, visual decision support in nutrition and offers design considerations for future proactive, user-adaptive health guidance interfaces.

Abstract

Seeking dietary guidance often requires navigating complex professional knowledge while accommodating individual health conditions. Knowledge Graphs (KGs) offer structured and interpretable nutritional information, whereas Large Language Models (LLMs) naturally facilitate conversational recommendation delivery. In this paper, we present HealthGenie, an interactive system that combines the strengths of LLMs and KGs to provide personalized dietary recommendations along with hierarchical information visualization for a quick and intuitive overview. Upon receiving a user query, HealthGenie performs query refinement and retrieves relevant information from a pre-built KG. The system then visualizes and highlights pertinent information, organized by defined categories, while offering detailed, explainable recommendation rationales. Users can further tailor these recommendations by adjusting preferences interactively. Our evaluation, comprising a within-subject comparative experiment and an open-ended discussion, demonstrates that HealthGenie effectively supports users in obtaining personalized dietary guidance based on their health conditions while reducing interaction effort and cognitive load. These findings highlight the potential of LLM-KG integration in supporting decision-making through explainable and visualized information. We examine the system's usefulness and effectiveness with an N=12 within-subject study and provide design considerations for future systems that integrate conversational LLM and KG.

Figures (6)

• Figure 1: Circular Interaction Workflow: Users query an LLM, which retrieves and visualizes relevant knowledge graph (subgraphs); users can then manipulate the graph to iteratively refine outputs. Demonstrated via HealthGenie, this cyclical loop enables adaptive, non-linear exploration of dietary recommendations.
• Figure 2: The overview of HealthGenie interface, which integrates a visualized nutritional knowledge graph and a conversational dialogue system powered by an LLM. Users can initiate interactions by asking nutrition-related questions and requesting dietary guidance, then LLM retrieves relevant information and generates informative responses (A). Users can explore more relevant information using query generation (B). Simultaneously, the corresponding nutritional information is visualized within a dynamic knowledge graph (C), allowing users to explore with more and less information (D). HealthGenie provides interaction tracing visualization, supporting users to perceive and operate their deletion or addition intuitively (E).
• Figure 3: Interaction with the Knowledge Graph: The visualized Knowledge Graph in HealthGenie enables users to explore a broader range of recipes (D). Users can interact directly with the visualized nodes, such as hovering to view detailed relationships or clicking to include or exclude ingredients (C) in the next recommendation. Any selection to include or exclude ingredients is recorded in the interaction history panel (E), ensuring all user actions are tracked.
• Figure 4: Participants' rating on output features of HealthGenie.
• Figure 5: Participants' responses regarding the task completion intuition, task completion accuracy and task completion satisfaction among four tasks, measured by the 5-point Likert scale questionnaire for both based and our system. Bars present the mean differences of our system compared to the Baseline. Dots indicated the 95% Confidence Interval.