Large language models bridge nutrition gap for cancer patients

Gemini stood out, offering visuals like photos and maps, plus exact store prices - a boon for patients pinching pennies. For a Latin American cuisine request, it even titled meals in Spanish, showing a knack for cultural nuance. ChatGPT, while less flashy, leaned into practical ingredient suggestions and store-brand tips.

Artificial intelligence (AI) could revolutionize dietary support for breast cancer patients, offering personalized meal plans to those sidelined by insurance gaps and limited access to nutrition experts. Researchers from Thomas Jefferson University’s Sidney Kimmel Medical College, in a paper titled "Bridging Gaps in Cancer Care: Utilizing Large Language Models for Accessible Dietary Recommendations," demonstrate that large language models (LLMs) like ChatGPT and Gemini can generate grocery lists and daily meal plans tailored to budget, culture, and location, potentially leveling the playing field for patients battling disparities in oncology care.

The study published in Nutrients underscores a stark reality: weight management, a proven factor in reducing cancer recurrence and boosting survival, remains out of reach for many.  Weight management is strongly linked to cancer prognosis and recurrence, particularly for hormonally driven cancers such as breast cancer. Yet despite its proven importance, nutritional counseling remains underutilized and inconsistently integrated into cancer care. Barriers include insurance limitations, limited availability of trained oncology dietitians, and logistical challenges, especially for underserved communities. Many patients are left to navigate dietary decisions alone, often without the knowledge, resources, or support necessary for success.

To evaluate whether AI tools could help bridge this gap, the research team developed 31 prompt templates designed to test the ability of ChatGPT (GPT-3.5) and Gemini (Version 1.0) to generate meal plans and grocery lists tailored to diverse patient needs. Variables included cancer stage, comorbidities, geographic location, culture, age, dietary guidelines, budget, and grocery store availability. The goal was to simulate real-world conditions in which patients might seek dietary advice adapted to their medical, financial, and cultural contexts.

Each prompt asked the LLMs to generate a daily meal plan and grocery list for a breast cancer patient. A subset of these prompts was also answered by board-certified oncology dietitians, allowing for comparison between human and AI-generated recommendations. The results were reviewed for nutritional adequacy, personalization, cultural sensitivity, and adherence to guidelines from the United States Department of Agriculture (USDA) and the Acceptable Macronutrient Distribution Range (AMDR).

The findings indicate that LLMs are capable of producing meal plans that are nutritionally balanced and culturally adaptable. While human dietitians were more precise in matching USDA-calculated daily calorie needs, the LLM-generated plans tended to better align with macronutrient distribution guidelines. Gemini demonstrated particular strength in adapting to budget constraints and cultural preferences, offering grocery lists with estimated prices and recipes in multiple languages. ChatGPT, meanwhile, focused more on meal preparation and practical suggestions for ingredient substitutions.

However, both AI models showed limited ability to account for individualized medical considerations such as comorbidities, treatment phase, and disease burden. Some recommendations—such as avoiding cruciferous vegetables in patients with COPD—were flagged by expert reviewers as overly general or potentially misleading. This underscores the current need for professional oversight and validation when applying AI-generated advice in clinical contexts.

Statistical comparisons revealed that while LLM and dietitian recommendations were broadly aligned, significant differences appeared in specific nutrient allocations. For example, the $100/day meal plan from Gemini included significantly more dietary fat than the dietitian’s version, while ChatGPT’s $10/day plan delivered more protein. Despite these variations, the overall nutritional profiles were largely similar, suggesting that LLMs can approximate expert-level dietary planning under certain conditions.

The study also highlighted key areas where AI models could improve. Neither model significantly varied its recommendations based on patient age, cancer stage, or comorbidity, suggesting limited responsiveness to these clinically relevant factors. Future iterations may benefit from fine-tuning or integrating clinical data to enable more precise personalization. Expanding the training data to include a wider variety of global cuisines and dietary traditions may also help reduce cultural bias, which researchers noted in the tendency of both models to default to American meal structures.

Despite these limitations, the accessibility of LLMs makes them a promising tool for delivering dietary support to patients who otherwise lack it. The ability to generate grocery lists, adapt to budget limitations, and respond in real time offers clear advantages, particularly for individuals navigating financial hardship, language barriers, or limited food access. As the burden on healthcare systems grows and demand for nutrition services outpaces supply, AI-driven solutions may offer scalable, cost-effective support.

Lastly, researchers stress that AI should not replace human dietitians but rather supplement their role, particularly in environments where professional nutrition support is unavailable or delayed. Integrating AI-generated plans into clinical workflows under expert supervision could streamline consultations and free up clinicians to focus on complex cases.