AI tool predicts hypoglycemia risk pre-exercise

A newly developed decision-support tool may help individuals with type 1 diabetes estimate their risk of hypoglycemia before exercise simply by using just three readily available inputs, according to a study published in Diabetologia.

"GlucoseGo represents an important advance in the prediction of exercise-associated hypoglycemia for people with type 1 diabetes," wrote Catherine L. Russon, PhD, and colleagues. "By distilling ML-[machine learning] derived predictions into a simple, user-friendly format, it supports informed decision-making and may promote physical activity and reduce hypoglycemia burden."

Practical Use

Fear of exercise-associated hypoglycemia remains a major barrier to physical activity among patients with type 1 diabetes, with only about half of adults with type 1 diabetes meeting recommended exercise targets, noted Dr. Russon of the University of Exeter Medical School, UK.

To address this issue, investigators developed and validated a simplified machine learning model designed to predict hypoglycemia risk at the start of exercise. The research combined data from 4 large studies involving 834 participants aged 12 to 80 and 16,430 recorded exercise sessions.

The team initially created a comprehensive predictive model incorporating 406 variables related to glucose levels, insulin use, and exercise characteristics. That model demonstrated strong predictive accuracy, achieving a receiver operating characteristic area under the curve (ROC AUC) of 0.89.

dHowever, the researchers then developed a simplified model designed for practical use outside research settings. By applying feature-selection methods, they found that three main variables—starting glucose level, exercise duration, and glucose trend arrows (determined as either “falling” or “stable/rising”)— could provide most of the predictive power. The simplified model achieved a comparable ROC AUC of 0.87, indicating only a modest decrease in performance to the initial model despite having dramatically reduced complexity.

The model was then translated into an accessible, user-friendly visual tool called GlucoseGo, a four-color traffic-light style heatmap that categorizes hypoglycemia risk as "very low," "low," "moderate," or "high" based on those three variables. In calibration analyses, observed hypoglycemia rates aligned closely with predicted risk categories, ranging from 0.2% in the very-low risk group to 43% in the high-risk group.

Across the dataset, hypoglycemia occurred in around 9.4% of exercise sessions, a rate observed as being consistent across all studies. Starting glucose level emerged as the strongest predictor of risk, while exercise duration and glucose trend further improved predictive accuracy.

Importantly, the simplified model maintained strong performance across independent validation cohorts and a variety of patient subgroups, including adolescents and individuals using multiple daily injections, insulin pumps, or closed-loop systems.

The developers of the tool noted that GlucoseGo is intended to complement, rather than replace, current exercise guidance for patients with type 1 diabetes. With further validation in more diverse populations and prospective clinical trials, the tool could help support safer exercise participation, reduce the psychological burden associated with hypoglycemia risk, and help integrate personalized risk assessment into routine diabetes self-management.

Study author John S. Pemberton reported his role as co-founder of EXercise in Type One Diabetes group, as well as advisory, research, and speaking relationships with multiple industry organizations, including Abbott, Roche, Dexcom, Insulet, Novo Nordisk, AstraZeneca, and Eli Lilly. The remaining authors reported no relevant conflicts of interest.

AACE Endocrine AI is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (AACE^®). The ideas and opinions expressed in AACE Endocrinology do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.