Abstract
AIMS/HYPOTHESIS: This study aimed to develop an accessible tool, derived using machine learning, to predict hypoglycaemia risk at the start of exercise and to provide clear, rapid risk assessment to support safer participation in exercise. METHODS: Data from four diverse studies were combined, encompassing 16,430 exercise sessions from 834 participants aged 12-80 years using various insulin delivery methods. The XGBoost algorithm was used to develop two models: a comprehensive model and a simplified model for predicting hypoglycaemia during exercise. RESULTS: The comprehensive model (406 variables) achieved a mean ROC AUC of 0.89. The simplified model, using only starting glucose, exercise duration and glucose trend arrows, achieved a comparable ROC AUC of 0.87. The simplified model performed consistently across exercise types and insulin delivery methods. In collaboration with individuals with type 1 diabetes, this model was translated into GlucoseGo, a user-friendly traffic-light heatmap displaying hypoglycaemia risk based on the three variables. CONCLUSIONS/INTERPRETATION: The GlucoseGo heatmap provides a practical, accessible tool for predicting hypoglycaemia risk immediately before exercise. It may empower individuals with type 1 diabetes to exercise more safely, reduce hypoglycaemic episodes, and increase engagement in physical activity.