Tool trained by AI can predict diabetes-related complications

Physicians may have a new tool to estimate a patient’s risk of developing diabetes-related complications. Researchers at the University of Maryland School of Medicine have developed and validated a new risk tool that can estimate the likelihood of a diabetes patient developing complications and update those estimates as new clinical information becomes available.

People with diabetes are at risk for heart disease, kidney disease, nerve damage, eye disease, and emergencies caused by very high or very low blood sugar. About 40 million people in the United States, or about 12% of the U.S. population, have diabetes, according to the CDC.

According to the CDC, there were about 16.5 million emergency department visits in 2021 among people with diabetes. A total of 8.1 million resulted in hospitalizations, with 1.78 million for major cardiovascular diseases. Of these, 230,000 were for a hyperglycemic crisis and 52,000 were for hypoglycemia; 166,000 results in the amputation of a lower extremity. In 2023, 12.3% of adults with diabetes reported severe vision difficulty or blindness.

While risk assessment tools exist, most focus on just one complication at a time or predict the risk of complications over a much longer period of time. The current models also usually rely on data from specialized research groups rather than real-world care settings.

Rozalina G. McCoy, M.D., MS, associate professor of Medicine in the Division of Endocrinology, Diabetes, and Nutrition, and her colleagues created a tool called the Diabetes Complications Risk Calculator (DCRC). This tool can estimate risk for nine different acute and chronic complications, including cardiovascular disease, stroke, kidney disease, nerve damage, and blood sugar crises.

“Our goal was to create a tool that reflects the reality clinicians face, where patients often have multiple concurrent and competing risks,” McCoy said in a news release. She is also director of the Precision Medicine and Population Health Program at the University of Maryland Institute for Health Computing. “By looking at these risks together and updating them over time, we can better understand what complication or complications our patients are most likely to experience, which can ultimately support more informed and actionable conversations between patients and their clinicians.”

McCoy and her team analyzed claims data from more than 400,000 adults newly diagnosed with diabetes Type 1 or Type 2 across the United States from the Optum Labs Data Warehouse, as well as data from electronic health records from Mayo Clinic for validation. They used machine learning to analyze insurance claims and electronic health record data. The models incorporate commonly available information such as age, existing health conditions, medications, and laboratory tests.

The Diabetes Complications Risk Calculator produces monthly risk estimates that change as a patient’s health status evolves. In testing, the models showed good to strong accuracy in predicting whether patients would develop specific complications, both in the original nationwide dataset and in an independent group of patients treated at Mayo Clinic.

In a paper published recently in Nature Communications, researchers said the tool has the potential to be applicable in real-world settings, but they said the generalizability of the tool with other patient cohorts will need to be tested. “We need to do more testing to understand how the tool performs when used in everyday clinical practice,” McCoy said.

Limitations of the tool include the use of data only from insured patients, which may not fully reflect patients without consistent access to medical care. In addition, some of its predictions were less accurate for certain complications.

The research team plans to evaluate how the calculator performs when integrated into real-world clinical workflows and whether it can improve shared decision-making and long-term health outcomes.