Loading ...

Artificial intelligence (AI) is steadily reshaping diabetes care, from diagnostic advances to providing new tools that lighten patients’ daily management burden, according to experts in the field.

“One of the biggest benefits of AI in diabetes management is the maturation of the hybrid closed-loop insulin pump and continuous glucose monitor systems,” said Yaa A. Kumah-Crystal, MD, associate professor of Biomedical Informatics and Pediatric Endocrinology at Vanderbilt University Medical Center in Nashville, Tennessee. As these systems evolve, she said, the goal is to automate more aspects of glycemic control and insulin dosing while giving users greater flexibility.

In a recent review in Endocrine Practice, Francisco Javier Pasquel, MD, of Emory University School of Medicine in Atlanta, and colleagues detailed where AI is delivering value now and where it’s headed.

“Clinically relevant AI is rapidly advancing across screening, risk prediction, decision support, and device optimization, with regulators and health systems beginning to adopt these tools,” Pasquel told Medscape Medical News. “Given this pace, clinicians need a current map of what’s real, what works, and what’s next.”

How AI Helps Today

Pasquel and colleagues noted that AI tools for diagnosing diabetic retinopathy and managing its associated complications are particularly primed for real-world application. US FDA-cleared AI systems now support automated retinal assessments that, in controlled settings, approach specialist-level diagnostic accuracy.

“This technology expands access to timely screening, particularly in primary care environments and in low-resource settings, including low-income countries, to address shortages of trained ophthalmologists,” Pasquel noted, while cautioning that many other potentially high-impact AI applications remain early in development.

Machine learning models that analyze thermograms and skin images are being studied to help pinpoint ischemia and infection in lesions associated with diabetic peripheral neuropathy. Predictive models are also improving risk stratification for incident diabetes, complications, hospitalizations, and hospital readmissions, helping clinicians target preventive care and resources.

Next-Generation Algorithms

Pasquel highlighted the promise of next-generation automated insulin delivery algorithms, which aim to detect and respond to unannounced meals and integrate contextual data (eg, exercise or sleep patterns) to optimize insulin delivery and reduce hypoglycemia risk.

He also pointed to a recent consensus statement by Peter G. Jacobs, PhD, of Oregon Health & Science University and colleagues, which outlines how machine learning is being embedded into sensors and drug delivery devices, identifying patterns that can improve individual health outcomes. However, the authors acknowledged that limited datasets and wide variability in glucose dynamics among individuals with diabetes make effective algorithm development challenging.

Data Synthesis and Patient Support

“Some of the most promising AI tools on the horizon allow us to synthesize the large amounts of data patients generate,” Kumah-Crystal told Medscape Medical News.

She gave the example of correlating a patient’s activity readings from their smartwatch with indicators such as heart rate to better gauge medication effects.

“Many of the tests that we have in endocrinology are invasive blood tests, but there are likely many other pieces of metadata, if patients would like to share them, that could be used with AI to build a more complete picture of their health,” Kumah-Crystal said.

As these tools mature, she envisions AI acting as a “coach and cheerleader,” providing timely insights and education and offering reminders about medications and appointments to engage patients in their own care.

Persistent Challenges and How to Meet Them

Kumah-Crystal noted that limitations currently surround the use of AI tools, especially in pediatrics and rare endocrine conditions.

“Many generative large language models may not have a clear understanding of the treatment and pathology of some of these rare conditions and are more likely to hallucinate and make up information about treatment strategies when that information is not in their training set,” she said.

This makes it imperative, she added, for academic institutions to vet consumer-facing AI tools and develop better patient education around potential shortcomings.

Pasquel pointed to additional hurdles, including inconsistent performance, variable data quality, biased datasets, and persistent concerns around privacy and interoperability. He called for standardized development and validation practices, adopting transparent and privacy-preserving approaches, seamless electronic health record integration, and ongoing human oversight.

“Regulatory processes, which often lag behind technological advances, will also need to adapt, incorporating more agile, risk-based review pathways to ensure that safe and effective tools reach clinical practice in a timely manner,” he emphasized.

Research Highlight: First Randomized Trial of AI to Reduce Diabetes Medications

Despite the challenges, evidence that AI can shift long-term outcomes for diabetes management is beginning to emerge. In a recent publication in NEJM Catalyst, Kevin Pantalone, DO, director of diabetes initiatives at the Cleveland Clinic, Cleveland, and colleagues reported a 52-week randomized trial of an AI-enabled program (Twin Precision Treatment) in adults with type 2 diabetes.

The intervention paired a smartphone app with four connected sensors (a continuous glucose monitor, an activity sensor, a smart scale, and a blood pressure meter), which delivered personalized recommendations about food choices, physical activity, and sleep, and included deep breathing exercises. The 100 patients in the intervention arm were compared with 50 patients receiving usual care.

At 12 months, 71% of participants in the AI intervention group achieved A1c < 6.5% without glucose-lowering medications other than metformin vs 2.4% with usual care. Furthermore, 22.6% of the intervention group reached that target with no glucose-lowering medications at all compared with 1.0% of the usual care group.

Overall, 88% of individuals randomized to using the AI intervention remained engaged with it for 12 months, which supports the favorability of the technology, and was reflected in improved treatment satisfaction and quality-of-life scores in the intervention group compared with the usual care group.

Pantalone told Medscape Medical News that he was struck by how readily patients across various ages and socioeconomic spectrums adopted the technology without difficulty, as well as just how effectively it performed.

“Not only did it improve blood sugar control and induce weight loss greater than usual care, but those benefits also occurred concurrently while the patients were stopping many of their glucose-lowering therapies, including GLP-1 receptor agonists.”

Putting It Into Practice

Pantalone cited access and adoption as the largest barriers to expanding AI in endocrinology. Although the system used in his study is commercially available and offered by some employers and health insurance plans, individuals cannot currently sign up independently to participate in a program or purchase the necessary technology at a retail location.

More randomized clinical trials, he added, are needed to support the benefits of AI-driven interventions in diabetes.

“Most of the benefits of AI-driven interventions for type 2 diabetes were based on companies’ real-world experience and analysis of their commercial data, which are not generally appreciated by healthcare providers, health plan directors, or third-party payers as reliable evidence,” Pantalone said. “Now that we have very clear evidence, from a randomized controlled trial, I believe the use of AI in the management of diabetes will become more common.”

Pasquel disclosed receiving research support from Insulet, Tandem Diabetes Care, Ideal Medical Technologies, Novo Nordisk, and Dexcom; personal fees from Dexcom; and consulting fees to Emory from Insulet.

Kumah-Crystal reported having no financial conflicts of interest.

Pantalone disclosed receiving institutional consulting and research support paid to Cleveland Clinic from Twin Health for the current study; consulting fees from Bayer, Boehringer Ingelheim, Corcept Therapeutics, Diasome, Eli Lilly and Company, Merck, Novo Nordisk, and Sanofi; speakers’ honoraria from AstraZeneca, Corcept Therapeutics, and Novo Nordisk; research support to Cleveland Clinic from Bayer, Novo Nordisk, and Twin Health; and a patent application.