Training AI to Speak Health: Diabetes Chatbot

Authors: Aaditya Malhotra (1), Bera Ankaya (2), Cindy Kuete (3), Gulsima Karsli (4), Marwa Hemat (5)

Editor: Fatma Zehra Tunca, George Washington University

Author’s Affiliations:  (1) Thomas Jefferson High School for Science and Technology, (2) Deep Run High School, (3) CMIT North High School, (4) George Mason University, (5) Alexandria City High School

Introduction

Did you know that 1 in 10 people in the United States live with diabetes? (CDC, 2024) Diabetes is a complex disease that causes disruptions in blood sugar regulation of the people. The disease involves many different types of treatments and life management. Therefore, to help manage their everyday lives, a personalized diabetes chatbot, Diabot, was developed.

Figure 1.0. Diabetes in the United States: A National Snapshot (CDC, 2024): This infographic presents key statistics about diabetes in the United States, including the number of affected individuals, the ratio of diagnosed cases, and the percentage of people unaware they have the condition.

Creating a reliable and personalized diabetes chatbot, Diabot, involves much more than just training it on clinical facts. For Diabot to truly help the people, it must reflect what daily life with diabetes actually feels like, which actually includes more than just numerical values or written symptoms. The data that is used must be able to capture the everyday experiences of the users in a realistic way. This could be things like how patients monitor their blood sugar, record their meals, adjust scheduled medications, describe what symptoms they are showing, and how they communicate with their healthcare providers.

While public health databases, academic studies, and trusted medical websites offer great resources, accessing high-quality data still remains quite difficult. In fact, some of the most meaningful information is either strictly protected by privacy laws, hidden within systems that do not communicate well with each other, or only made available through complex research partnerships. Even though these protections are essential for safeguarding patient confidentiality, they do also make it harder to collect the kind of rich and comprehensive data that projects like Diabot need to learn from reliable and effective data.

Figure 2.0. Key Use-Cases for Healthcare Chatbots: This figure outlines five primary applications of healthcare chatbots, including symptom analysis, appointment scheduling, mental health support, query management, and post-treatment care, emphasizing their role in streamlining patient services and improving accessibility.

Making the Data Work

Nevertheless, even if useful information is able to be collected, it often will come in many pieces. This is especially because the records may leave out certain details, use unfamiliar or medical wording, or follow inconsistent formats. Before the Diabot can learn from this provided information, it will have to be cleaned, standardized, and double-checked. A phrase like “low sugar” might appear in one record, while another uses “hypoglycemia.” These must be trained to be treated as the same thing.

It is important to note that if the model is trained on messy or mismatched data, it could actually pick up the wrong information. This could lead to responses that are not just unhelpful, but also potentially risky or harmful. Therefore, taking the time to prepare the accurate data in the right way means users will get answers they can actually trust.

 

Combining GPT with Expert Guidance

Artificial intelligence has already been reshaping many industries, including healthcare. Certain tools like GPT have been proven to offer incredible potential. Only when trained carefully, these models can provide diabetic patients with thoughtful and personalized support. They can do numerous tasks, such as answering questions, making appropriate suggestions, and helping people stay consistent with their care and routine to keep things stabilized.

Of course, GPT is not perfect. That is why Diabot is designed to work alongside content that has already been reviewed by medical professionals. In fact, Diabot is fine-tuned to reflect specifically the prepared and trained information. This prevents the kind of hallucinations that AI models sometimes produce. For example, if a user asks, “What medication should I take for type 2 diabetes?” Diabot does not guess. Instead, it points the user to credible resources from organizations like the ADA, CDC, NIH, or ADCES.

In addition, Diabot also offers help with meal planning, tracking symptoms, understanding medical language, and sticking to treatment goals. It speaks in a way that is easier to understand, while still being medically grounded.

 

Setting Boundaries for Safety

Diabot is also built within certain rules that help it avoid any unsafe territory. The especially trained rules guide the chatbot’s responses when a user brings up something serious. For example, if a user says, “I have chest pain,” Diabot does not attempt to diagnose or offer advice. Instead, it responds immediately with a warning to seek emergency care services. It is not meant to be a doctor, instead, the Diabot is there to guide patients through the guidance they receive from their professional healthcare providers. This kind of built-in awareness helps protect users from putting their trust in a tool that is not meant to replace professional medical judgment.

 

How Accurate Are the Responses?

To see how reliable AI-generated medical advice can be, many types of research have been implemented over the past couple of years. According to a report in National Institute of Health (2024), across 284 questions developed by 33 physicians in 17 different specialties, a chatbot’s median accuracy score was 5.5 out of 6. The answers were found to be mostly accurate and helpful, although there were still some limitations (National Institute of Health, 2024).

As mentioned before, the Diabot is not supposed to be a replacement for doctors. It is a support tool for patients living with diabetes. And as more doctors begin using AI in their work - with 66 percent already doing so in 2024 (American Medical Association, 2025) - it actually makes sense that patients could benefit too. Having a chatbot like the Diabot would make managing diabetes feel a little less overwhelming.

 

Avoiding AI Hallucinations

One of the biggest challenges in chatbot design is reducing AI hallucinations. These, what they call hallucinations, actually happen when an AI model gives an answer that sounds very confident but is not actually true. In medicine, that is not just a problem, but it can also be quite dangerous. Therefore, Diabot tackles this by using something called retrieval-augmented generation. Instead of guessing or filling in gaps, it pulls information from reliable sources that it has been trained with, before responding. For example, if a user asks about an unusual side effect, Diabot either gives the previously researched and trained trustworthy answer or instead recommends speaking directly with the healthcare provider. In order to protect the users and build trust, it does not invent answers on its own.

 

Understanding the Human Behind the Question

Diabot is designed to do more than just provide medical facts. It also listens for and acknowledges the emotion of the users, and the urgency of their prompt. There is a difference between someone asking, “What should I eat today?” and someone asking, “Am I going to get worse?”

By using natural language processing (NLP), Diabot is able to recognize tone and intent of the users. If, for example, someone says, “I feel dizzy and my sugar is high,” Diabot understands that this might be an emergency as it involves a serious side effect. It responds gently but also immediately recommends professional help. On the other hand, if someone casually asks about a food choice, Diabot offers clear guidance and an informative explanation.

 

Testing with Real Healthcare Experts

Before Diabot ever reaches a real user, it goes through rounds of testing with healthcare professionals. Doctors and other healthcare professionals read through Diabot’s answers and give feedback about whether they are too long, too vague, or missing something important.

The development team also runs sample conversations to see how Diabot performs. These include everything from insulin reminders to blood sugar management and nutrition. Doctors, nurses, and diabetes educators help with preparing the fine-tune the responses so they are medically sound and easy to understand. For example, the consulting doctor might say, “This is a great explanation, but it is too much text for someone reading on their phone.” That kind of input actually helps the team make responses more concise without losing the important message.

 

Conclusion

At its core, Diabot is built to support patients living with diabetes, not replace anyone in the healthcare system. It is here to give patients more confidence and clarity about their conditions as they manage a complicated condition.

By using reliable information, listening for emotional needs, checking its own limits, and working with real experts, Diabot shows how AI can be helpful in healthcare not only as a substitute for human care, but also as a companion that helps people feel less alone and more in control.

Glossary

1. AI (Artificial Intelligence): The ability of machines or computer programs to perform tasks that typically require human intelligence, such as learning, understanding language, reasoning, and problem-solving.

2. ADA (American Diabetes Association): A nonprofit organization that focuses on diabetes research, education, and advocacy. It provides widely accepted clinical guidelines and resources for people living with diabetes.

3. ADCES (Association of Diabetes Care & Education Specialists): A professional organization that supports diabetes educators and promotes evidence-based care practices for patients managing diabetes.

4. Anonymization: The process of removing or masking personally identifiable information from health data so that individuals cannot be identified, helping to protect patient privacy in AI development and medical research.

5. CDC (Centers for Disease Control and Prevention): A U.S. federal agency responsible for protecting public health through research, education, and the prevention of disease, including diabetes.

6. Diabot: A personalized, AI-powered chatbot designed to help people manage diabetes by offering medically informed guidance, symptom tracking, and emotional support. Diabot is trained with medically reviewed content and built to recognize urgency and tone in user input.

7. Endocrinologist: A medical doctor who specializes in treating hormone-related conditions, including diabetes and other disorders involving the endocrine system.

8. GPT (Generative Pre-trained Transformer): A type of artificial intelligence language model that can understand and generate human-like text. Diabot uses a version of GPT fine-tuned with medical information and expert-reviewed content.

9. Hallucination (AI): When an AI model produces information that sounds convincing but is actually false or not based on verified sources. This is especially dangerous in healthcare settings.

10. HbA1c: A blood test that measures the average blood glucose levels over the past 2 to 3 months. It is used by doctors to diagnose diabetes and track how well a person is managing the condition.

11. HIPAA (Health Insurance Portability and Accountability Act): A U.S. law that ensures the privacy and security of patients’ health information and medical records.

12. Hypoglycemia: A medical condition where blood sugar levels drop below normal. This can be dangerous and often requires immediate treatment, such as consuming glucose.

13. Insulin: A hormone produced by the pancreas that allows the body to absorb and use glucose for energy. People with diabetes may need insulin therapy if their bodies do not produce or use insulin effectively.

14. National Institutes of Health (NIH): The primary medical research agency of the U.S. government. It supports scientific studies, including those on diabetes and artificial intelligence in healthcare.

15. Natural Language Processing (NLP): A field of artificial intelligence that enables machines to understand and interpret human language. In Diabot, NLP is used to detect user tone, urgency, and the intent behind questions.

16. PHI (Protected Health Information): Any health-related data that can be linked to an individual. Under HIPAA, PHI must be kept secure and confidential, especially in AI training and deployment.

17. Retrieval-Augmented Generation (RAG): An AI technique that improves answer accuracy by retrieving facts from external sources before generating a response. Diabot uses this method to avoid hallucinations and ensure its answers are grounded in reliable medical information.

18. Rule-Based System: A set of pre-programmed rules built into an AI model to prevent it from making unsafe or irresponsible decisions. In Diabot, these rules are used to handle emergency prompts and sensitive situations properly.

References

David Priede, MIS. “Are Healthcare AI Models Trained with My Health Data? By David L. Priede, MIS, Phd.” Biolife Health, Biolife Health, 23 Apr. 2025, www.biolifehealthcenter.com/post/are-healthcare-ai-models-trained-with-my-health-data.

Healthcare Chatbots - Their Benefits, Use Cases, and Examples. www.openxcell.com/blog/healthcare-chatbots/. Accessed 17 July 2025.

Henry, Tanya Albert. “2 in 3 Physicians Are Using Health AI—up 78% from 2023.” American Medical Association, 26 Feb. 2025, www.ama-assn.org/practice-management/digital-health/2-3-physicians-are-using-health-ai-78-2023.

Johnson D, Goodman R, Patrinely J, et al. “Assessing the Accuracy and Reliability of AI-Generated Medical Responses: An Evaluation of the Chat-GPT Model.” Research Square, 2023 Feb:rs.3.rs-2566942. DOI: 10.21203/rs.3.rs-2566942/v1. PMID: 36909565; PMCID: PMC10002821.

Javaid, M., Haleem, A., & Singh, R. P. (2023). “ChatGPT for Healthcare Services: An Emerging Stage for an Innovative Perspective.” BENCHCOUNCIL Transactions on Benchmarks, Standards and Evaluations, 3, Article ID 100105. www.scirp.org/reference/referencespapers?referenceid=3793783. Accessed 16 July 2025.

“A Report Card: Diabetes in the United States Infographic.” Centers for Disease Control and Prevention, www.cdc.gov/diabetes/communication-resources/diabetes-statistics.html. Accessed 16 July 2025.

Nandigrami P, Szczepaniak F, Boughter CT, Dehez F, Chipot C, Roux B. “Computational Assessment of Protein-Protein Binding Specificity within a Family of Synaptic Surface Receptors.” The Journal of Physical Chemistry. B, U.S. National Library of Medicine, pubmed.ncbi.nlm.nih.gov/35787023/. Accessed 16 July 2025.

“The Importance of Data Anonymization in Ensuring HIPAA Compliance for Healthcare AI Applications - Simbo AI - Blogs.” Simbo AI - Blogs, www.simbo.ai/blog/the-importance-of-data-anonymization-in-ensuring-hipaa-compliance-for-healthcare-ai-applications-426075. Accessed 16 July 2025.

“Use of Small-Area Estimates to Describe County-Level Geographic Variation in Prevalence of Extreme Obesity among US Adults | Nutrition, Obesity, Exercise | JAMA Network Open | JAMA Network.” jamanetwork.com/journals/jamanetworkopen/fullarticle/2765684. Accessed 17 July 2025.

Wang, Yingtan, et al. “Pathological Gambling in a Patient on Piribedil: A Case Report.” Medicine, U.S. National Library of Medicine, 12 Feb. 2021, pmc.ncbi.nlm.nih.gov/articles/PMC7886474/.

“AI for Diabetes Management: Personalized Care and Early Detection.” Newswise, 25 Feb. 2025, www.newswise.com/articles/ai-for-diabetes-management-personalized-care-and-early-detection.