AI Reshaping Diabetes Care From Nutrition to Clinical Decisions

Artificial intelligence (AI) is increasingly moving from research promise to practical application in diabetes care, offering tools that address persistent challenges. Presentations at the American Diabetes Association 2026 Scientific Sessions, in the symposium, “AI Integration Into Diabetes Technology,” examined 2 of these challenges: the burden of nutrition estimation on patients and the limits of real-time clinical decision-making.

Underscoring the presentations was a shared tension between algorithmic capability and the irreplaceable role of human judgement. Researchers tackled these challenges from distinct angles. One focused on AI-powered tools to simplify carbohydrate counting for patients managing type 1 diabetes (T1D), and the other focused on decision support systems that are approaching, though not yet replacing, the clinical reasoning of experienced endocrinologists.

AI Takes the Guesswork Out of Carb Counting¹

For the approximately 2.1 million Americans living with T1D² who together account for 5% to 10% of all diabetes cases,³ calculating the carbohydrate content of every meal is not a minor inconvenience. It is a daily clinical task with real consequences for glycemic control. Yao Qin, PhD, MS, assistant professor at University of California, Santa Barbara Robert Mehraan College of Engineering and AI researcher at Google Brain, knows this firsthand. Qin, who manages T1D herself, described the manual carb counting process as time consuming and prone to error, particularly for patients without dietician training, in her presentation, “AI for Automatic Nutrition Estimation.”

“Whenever I take a meal, I take a pause and I start a Google search of how many carbohydrates in each food item, and I need to search 1 by 1,” she said. “At the end, I also need to add all of them together, do simple math. But when the items are quite high, then it still takes time. So this definitely makes our lives much harder. In reality, what I really do is skip all this process and just do a random guess.”

Her team’s response is NutriBench, an AI-powered nutrition estimation tool that allows users to describe a meal in natural language and receive a detailed carbohydrate breakdown. The system supports multiple languages, can integrate with Apple Health, and is accessible via WhatsApp and mobile devices—design choices, Qin explained, that reflect an explicit focus on global usability.

To validate the tool’s accuracy, Qin and her team built a dataset of over 15,000 meals, spanning 24 countries and benchmarked the AI against human dieticians. The AI outperformed dieticians on speed and matched or exceeded their accuracy, particularly when dieticians lacked access to supplemental nutrition references. A data science simulation was used to model how each group’s carbohydrate estimates translated into predicted blood glucose outcomes, a clinically meaningful test of real-world performance.

To fine-tune reliability, the tool also incorporates reinforcement learning. Along with each estimate, the model outputs a confidence score, flagging uncertainty rather than generating false precision. Qin emphasized this as essential for safe integration into insulin dosing decisions. Accurate, real-time nutrition data has implications across obesity, cardiovascular disease, cancer, and neurological health, condition where diet plays a causal or modifying role but where patients and clinicians alike lack practical tools for tracking intake and the level of detail AI now makes possible.

When AI Thinks Like a Doctor, and When It Doesn't

In some clinical tasks, AI tools in diabetes are no longer merely assisting physicians. They are performing comparably to them. Revital Nimri, MD, director of the scientific and technology diabetes service at Shneider Children’s Medical Center of Israel, made this case today in her presentation, “AI Decision Support Tools: Ready to Replace MDs?” She presented clinical trial data showing that AI-generated insulin dose recommendations produced glycemic control outcomes comparable to those of experienced endocrinologists.

“AI can collect, integrate, analyze, predict, and support decisions across many clinical tasks. Some are already being taken over by AI, and many others are in development,’ she said. “This includes identifying individuals at risk for diabetes onset or complications and interpreting [continuous glucose monitor] data.” For example, machine learning prediction for hypoglycemia and hyperglycemia, insulin titration, and treatment selection.

In one 6-month randomized trial, AI-generated insulin titration recommendations were compared head-to-head against recommendations from physicians, with comparable glycemic outcomes. Clinician acceptance of AI’s recommendations ran as high as 97%. Another prospective analysis that comprised several studies, Nimri explained, incorporated a stepwise approach to AI-driven insulin adjustment, gradually reducing physician oversight while monitoring for safety signals via automatic safeguards. The results demonstrated significant improvements in glycemic control alongside appropriate insulin dose adjustments.

Nimri then introduced a more expansive application, Agent AI, a software support system designed to support the diabetes care process from end to end. The system operates continuously, adapting its recommendations based on physician feedback through a learning loop. She also discussed a prototype concept, a digital twin agent for endocrinologists, that is currently being tested to improve glycemic control by modeling individual patient responses over time.

She was equally direct about AI’s limits. It is constrained by the data it is trained on, and tools validated in one clinical or demographic context may underperform in another. She also raised the issue of accountability: AI can generate accurate recommendations, but it cannot assume clinical responsibility, and it lacks the empathy and relational dimensions often associated with better patient outcomes.

The future Nimri described is not one of replacement but calibration: deploying AI where it demonstrably performs, preserving human oversight where it matters most, and continuing to interrogate questions of equity, trust, and responsibility as these tools move deeper into routine care.

References

1. Qin Y. AI for automatic nutrition estimation. Presented at: American Diabetes Association 2026 Scientific Sessions; June 5-8, 2026; New Orleans, LA.
2. National Diabetes Statistics Report. CDC. January 21, 2026. Accessed June 6, 2026. https://www.cdc.gov/diabetes/php/data-research/index.html
3. Diabetes basics. CDC. January 2, 2026. Accessed June 6, 2026. https://www.cdc.gov/diabetes/about/index.html
4. Nimri R. AI decision support tools: ready to replace MDs. Presented at: American Diabetes Association 2026 Scientific Sessions; June 5-8, 2026; New Orleans, LA.
5. Evaluation of the Dreamed Advisor Pro for automated insulin pump setting adjustments in children and adolescents with type 1 Diabetes- the Advice4U Pro Study (Advice4U). ClinicalTrials.gov. Updated December 27, 2019. Accessed June 6, 2026. https://clinicaltrials.gov/study/NCT03003806