Identification of Activity, Insulin, and Dietary Thresholds Associated With Exercise-Related Hypoglycaemia and Hyperglycaemia in Children With Type 1 Diabetes: A Free-Living Study Using Explainable Machine Learning.
Iqbal A, Dahimi H, Preux P, Lespagnol E, Melin A, Morel P et al. · Diabetes, obesity & metabolism · 2026
Aims: Some studies have explored associations between physical activity (PA) and hypoglycaemia in real-life in type 1 diabetes (T1D) but without fully accounting for two major confounders, diet and insulin. We aimed to identify thresholds of PA characteristics, insulin, and carbohydrates associated with dysglycaemia across the exercise-recovery cycle in children with T1D.
Materials and methods: Continuous glucose monitoring and accelerometry data, self-reported PA sessions (timing, duration and perceived intensity), diet (timing, type and quantity; optional photographs) and insulin data (doses and timing; corrective/meal boluses and basal insulin) were collected in 36 children with T1D (11.9 ± 3.3 years, injections or open-loop pumps) over seven free-living days. Accelerometer data were analysed for periods corresponding to self-reported PA sessions. Ensemble machine-learning models classified hypoglycaemia ( 180 mg/dL) during three phases: PA, 2-h post-exercise (early recovery), overnight. Shapley analysis identified risk and protection thresholds of features with high importance.
Results: Models achieved moderate to strong performance (AUC: 0.65-0.99; F1-score: 0.62-0.95) across outcomes and phases. Insulin boluses > 11% of total daily dose within 4-h pre-exercise and > 17% during early recovery were associated with hypoglycaemia risk during PA and early recovery respectively. Carbohydrate intake showed collinearity with insulin, resulting in complex associations with glycaemia. Self-reported PA > 80 min was associated with hypoglycaemia risk during PA, while accumulating > 15 min of accelerometer-derived vigorous PA was actually protective against early recovery hypoglycaemia. Multiple daily sessions were associated with nocturnal hyperglycaemia protection.
Conclusions: Phase-specific thresholds across exercise and insulin domains associated with exercise-related dysglycaemia were identified.