Health Information Dashboard: a large language model-based chatbot at the Jacobs Center for Precision Nutrition and Health

Large language models (LLMs) can facilitate research accessibility to disseminate over the ‘last mile’ for individuals who can directly apply them to their work or lives. At the Jacobs Center for Precision Nutrition and Health, we have developed a set of platforms for users looking for reputable health information. The pain point we sought to address was that health information online may not always be evidence-based, trustworthy or current. Using generative AI restricted to only peer-reviewed, evidence-based sources (Campbell Collaboration, Cochrane Collaboration, NIH Office of Dietary Supplements, Wikipedia, and WHO), our Health Information Chatbot answers queries such as “tell me about fortification for anemia” and provides citations for further reading. If the question is asked in another language, the chatbot will translate source material and respond in the same language (currently, we include 12 of the most commonly spoken language/dialects). Guardrails against hallucination, a common problem in generative AI, were also put into place; in other words, a question that asks about information not contained in the restricted data sources will be answered with “No relevant information” rather than a made-up (“hallucinated”) answer. Once an LLM-based chatbot responds to a query, the information is contained in the same chatbot interface – basically, a response message/bubble. However, sometimes it is more useful to have this information output as a Fact Sheet, that can be easily referred to for future use. We created second version of our Health Information Chatbot, Health Information Fact Sheet Chatbot, that does just that. Our factsheet includes headers on: Problem, Current WHO Recommendations, Current State of the Evidence, and Ongoing Clinical Trials for a well-rounded set of information related to the query. User-friendly, AI-based chatbot interfaces represent a useful, modern approach to democratizing data access and distillation for multiple audiences/user-groups, including students, researchers, consumers, and policymakers.