Study Explores Value and Shortcomings of AI-Enhanced Wearable Devices for Type 2 Diabetes and Prediabetes Care

University at Buffalo researchers have published in NPJ Digital Medicine the first comprehensive meta-review (a study of studies) of AI-enhanced wearables for people with prediabetes and Type 2 diabetes. Their conclusion is that these devices have enormous potential that will be realized once certain challenges are overcome.

Just how much AI-enhanced wearable devices could help patients with diabetes became clear to Raphael Fraser, PhD, corresponding author and research associate professor of medicine at the Jacobs School of Medicine and Biomedical Sciences at UB, when he saw that AI-enhanced CGMs can provide data every few minutes, instead of just a few readings per day.

“As newer AI models demonstrated the ability to recognize patterns and predict glucose changes before they happened, it became clear that diabetes care could shift from reacting to problems after they occur to anticipating and preventing them,” Fraser says. “That was the moment when I realized AI could genuinely transform daily management and long-term outcomes. AI turns CGMs from a rear-view mirror into a heads-up display.”

“For people living with diabetes, AI-enabled wearables have the potential to provide more timely and personalized guidance, helping them avoid glucose swings and manage daily decisions with greater confidence,” says Fraser. “For clinicians, the key takeaway is that these tools may help identify risks earlier and support more efficient care.”

While larger studies are needed to evaluate the benefit of CGMs for people with prediabetes, early use of wearables combined with AI could support lifestyle changes and perhaps delay or prevent the progression to diabetes.

Research in this area is growing rapidly, but the studies done to date have been scattered across different devices, data types and AI models, making it difficult to see the bigger picture.

“We wanted to bring everything together to understand what we actually know, what has been consistently shown and where the evidence is still thin,” explains Fraser. “Our goal was to identify which approaches seem most effective, where the limitations are and what gaps need to be addressed before AI-enabled wearables can become routine tools in clinical care.”

The researchers, all in the Division of Population Health in the Department of Medicine at the Jacobs School, selected 60 out of 5,000 peer-reviewed studies that examined the integration of AI and wearable technology in diabetes management.

There were many positive findings. “AI-enhanced wearables can predict glucose changes up to one to two hours in advance, helping individuals maintain steadier control and receive personalized guidance that reflects their daily routines, activity levels and sleep patterns,” Fraser says.

These systems also have the potential to reduce clinical workload by sorting through large streams of data and highlighting what requires attention.

But the researchers also found aspects of AI-enhanced wearables that were problematic; for example, AI-enhanced wearables are based on different AI models, which must be transparent and validated before they are widely adopted, Fraser says.

“Many AI models operate as ‘black boxes,’ making it difficult for clinicians and patients to understand or trust their recommendations,” Fraser explains, “which limits their usefulness in guiding day-to-day decision-making for people with prediabetes and Type 2 diabetes.”

For example, he says, an AI-enabled glucose app may warn a user that their blood sugar is likely to rise in the next 30 minutes but provide no insight into what triggered the prediction, whether that is a recent meal, reduced physical activity, elevated stress, poor sleep or normal day-to-day fluctuations.

“When people cannot see the ‘why’ behind the alert, they struggle to decide what action to take, making the tool far less helpful in real life,” says Fraser.

He adds that in some of the studies, limited sample sizes and narrow demographic representation reduce how broadly the findings can be applied. In addition, the lack of standardized benchmark datasets means results are not always easy to compare across studies. Practical barriers, such as inconsistencies in the quality of the data included in the studies, limited integration into clinical workflows, and the cost and accessibility of wearable devices, also curb widespread adoption of these devices, says Fraser.

Another factor that influences how well a CGM will work for patients and clinicians is what type of AI model it uses.

“Different AI models are suited to different kinds of data and prediction tasks,” explains Fraser. “Models designed to learn patterns over time, such as long short-term memory networks or similar architectures, tend to perform better with continuous glucose data because they can track trends and anticipate future changes,” he says. “Newer models like transformers are particularly good at integrating multiple forms of data, such as glucose, heart rate, sleep and physical activity, which can give a more holistic understanding of the body’s metabolic state.”

Nevertheless, he adds, sometimes simpler models are easier for clinicians to interpret. “So the challenge is not just choosing the most powerful model but choosing one that performs well while also being understandable and clinically trustworthy,” he says. “The ‘right’ AI is the one that fits the data and can be explained in the doctor’s office.”

Co-authors from the Jacobs School were:

• Rebekah J. Walker, PhD, associate professor of medicine
• Jennifer A. Campbell, PhD, MPH, associate professor of medicine
• Obinna Ekwunife, PhD, assistant professor of medicine
• Leonard E. Egede, MD, Charles and Mary Bauer Endowed Chair and professor of medicine

The research was supported by the American Diabetes Association, the National Institute of Diabetes and Digestive Kidney Disease, and the National Institute for Minority Health and Health Disparities.