Research shows that good nutrition can contribute to the overall health and well-being of cancer patients. However, personalized dietary advice can be hard to access and is not often covered by health insurance.
A research team from Thomas Jefferson University wondered if artificial intelligence (AI) bots might be able to close this gap in supportive care. In a new paper , they tested two widely available large language models (LLMs)—ChatGPT and Gemini— with a variety of prompts to generate meal plans and grocery lists for specific patient characteristics such as cancer stage and other health conditions, as well as budgets, geographic location, and cultural preferences.
The goal, says senior author and radiation oncologist at Sidney Kimmel Comprehensive Cancer Center Nicole Simone, MD , was to evaluate the LLMs capacity to address these varied needs, and compare the generated advice to that of clinical dietitians.
Medical student and first author, Julia Logan, tested the prompts for all the variables during her summer research stint in the department.
Co-author and computational physicist Wookjin Choi, PhD says it would be nearly impossible to build a specialized model from scratch. "The general LLMs provide a starting place, and then we develop it, score it, adjust it and create a new specialized model," says Dr. Choi, who is a member of Sidney Kimmel Medical College.
The researchers were pleasantly surprised by the level of granularity the tools achieved, Dr. Simone adds. "It was able to figure out cost and nearby grocery stores with available ingredients."
"The ability of LLMs to generate meal plans tailored to socioeconomic and cultural considerations is one of the most promising aspects of this approach," the authors write. That means the tool can be helpful for people with financial constraints and promotes adherence by integrating ethnic food choices.
This study demonstrated the feasibility of creating an AI tool to generate and support dietary advice tailored to cancer patients. There's more work to be done, such as identifying the limits of an AI tool's reliability and when oversight by professional dietitians is needed.
By Jill Adams
