Skip to main content
NUTRITION 2026 Main banner

(PTFS06-05-26) Evaluating the Reliability of Large Language Models for Food Micronutrient Inference From Dietary Images

Sunday, July 26, 2026
11:16 AM - 11:20 AM ET
Location: Prince George's Hall B-E (The Hub) - Poster Theater 2 / Poster Board Number: PT2-5
Nutrition Translation Award Competition for Early Career Investigators

    Presenting Author(s)

  • Runze Yan, PhD

    Assistant Professor
    Emory University
    Atlanta, Georgia, United States

    Disclosure(s): No relevant financial relationship(s) with ineligible companies to disclose.

    • Co-Author(s)

  • JL

    Jiaying Lu, PhD

    Assistant Research Professor
    Emory University
    Atlanta, Georgia, United States

  • XH

    Xiao Hu, PhD

    Emory University, United States

  • Hanqi Luo, PhD

    Assistant Research Professor
    Emory University
    Oakland, Georgia, United States

Disclosure(s):
Runze Yan, PhD: No relevant financial relationship(s) with ineligible companies to disclose.
Objectives: Large language models (LLMs) are AI systems increasingly used for food image analysis and dietary assessment, but their reliability in estimating micronutrient content remains unvalidated. Given that micronutrient intake is closely linked to chronic disease prevention, inaccurate estimates could mislead dietary guidance. This study evaluated five LLMs in estimating 65 nutrients from food images.

Methods: We tested three commercial models (GPT-5, Gemini-3, Claude Sonnet 4.5) and two open-source models (Qwen3-VL-8B, GLM-4.6V-Flash) on two public datasets with ground-truth labels for 65 nutrients: ASA24 (3,000 standardized food images with 3 portion sizes each) and SNAPMe (5,000 food images from 95 participants across 275 study days). Each model was prompted to identify food items, estimate portion sizes, and report nutrient amounts, and could decline when uncertain. Performance was evaluated on: (1) response completeness, measured by non-response rate (declined estimates) and hallucination rate (implausible extreme values), and (2) estimation accuracy, quantified using a range-normalized error metric for fair comparison across nutrients.

Results: Commercial models showed higher response completeness (mean non-response and hallucination rate: 16.2%) than open-source models (39.5%). Claude Sonnet 4.5 had the lowest rate (7.5%); GLM-4.6V-Flash had the highest (62.1%). All models performed worse on micronutrients than macronutrients (e.g., GPT-5: 6% vs. 29%). Although hallucination was generally rare, specific nutrients were more susceptible, with added vitamin B12 reaching 1.2% in Gemini-3 and certain polyunsaturated fatty acids reaching 0.4% in GPT-5. For estimation accuracy, even the best model (Claude Sonnet 4.5) had a mean normalized error of 116.3%, with over 78% of nutrients exceeding 100% error. Commercial models outperformed open-source models (125.7% vs. 149.1%).

Conclusions: All five LLMs showed substantial limitations in micronutrient inference, with high non-response rates and poor accuracy. These findings suggest that current LLM-generated nutrient estimates are not yet reliable for dietary assessment or clinical use. Domain-specific fine-tuning and integration with authoritative nutrient databases may help address these limitations.

Funding Sources: N/A