A Machine-Learning Framework based on Continuous Glucose Monitoring to Prevent the Occurrence of Exercise-Induced Hypoglycemia in Children with Type 1 Diabetes

Physical activity is recommended in patients with type 1 diabetes (T1D), but therapy management still lacks efficient tools to avoid exercise-induced hypoglycemia. Machine learning represents a powerful solution in the field of decision support for diabetes management and its application to continuous glucose monitoring (CGM) data appears promising in pre-exercise prediction of upcoming adverse events. Aim of this study was to investigate the possibility to distinguish if a specific configuration of CGM metrics evaluated before starting of exercise is more prone to induce hypoglycemia after the start of the exercise session until the following day. A total of 47 CGM recordings from T1D children have been used to extract CGM metrics from pre-exercise CGM data. Acquisitions were labelled as HYPO or as NO-HYPO, respectively if belonging to subjects who experienced or did not experience hypoglycemia during the time following the exercise. Anthropometric characteristics and extracted features have been given as input to a decision tree classification algorithm to select those with the most predictive power. The selected features were then further evaluated with respect to the classification problem by using them as input to other three classification models: random forest, adaboost and gradient boosting. Performance results in terms of area under receiver operating characteristic (AUC) were as follows: 85.5%, 82.1%, 78.1% and 74.3% for decision tree, gradient boosting, random forest and adaboost, respectively. M-value, maximum glucose, time above 180 mg/dL and time above 250 mg/dL could have a role in predicting upcoming hypoglycemia prior the starting of exercise.