The datasets correspond to all data used in the analyses of the paper "Explainable machine learning to identify activity-related, insulin and dietary thresholds for prevention of exercise-associated hypoglycaemia and hyperglycaemia in free-living children with type 1 diabetes". All information about the data collection are provided in the paper. Please find below a summary of the study and the results obtained with the data.

Physical activity induces immediate and delayed glycaemic changes in children with type 1 diabetes (T1D), yet real-life thresholds for exercise characteristics that jointly account for insulin and diet remain poorly defined. In a 7-day free-living study of 36 children with T1D, we combined continuous glucose monitoring with self-reported and accelerometer-derived exercise measures, alongside detailed insulin and dietary data. Ensemble machine-learning models predicted hypoglycaemia (180 mg·dL⁻¹) during exercise, the 2-hour post-exercise recovery, and the subsequent night, with thresholds identified using Shapley value analysis. The models achieved high predictive accuracy and revealed distinct phase-specific patterns. Pre-exercise glucose levels >175 mg·dL⁻¹ were associated with glucose decline during activity, whereas levels >150 mg·dL⁻¹ predicted glucose increases during early recovery. Insulin boluses exceeding 11% of the total daily dose in the 4 h before exercise and 17% during the 2 h recovery period increased hypoglycaemia risk. Exercise effects were phase dependent: self-reported activity lasting >80 min increased hypoglycaemia risk during exercise, while >15 min of objectively measured vigorous activity reduced risk during early recovery. A pre-sleep glucose range of 100–115 mg·dL⁻¹ minimised nocturnal dysglycaemia, and behavioural patterns such as afternoon exercise and multiple daily sessions were associated with lower overnight hyperglycaemia. Patient characteristics, including age, BMI Z-score, and daily insulin dose, influenced glycaemic risk in phase-specific and sometimes opposing ways. Overall, this analysis identifies quantitative, phase-specific thresholds for exercise and insulin management across the exercise–recovery cycle, moving beyond arbitrary categorisation used in previous studies.