Machine Learning-based Approach for Predicting Exercise-Induced Nocturnal Hypoglycemia in Type 1 Diabetes Patients Using Physiological Features, Temporal Dynamics, and Data Augmentation

Nocturnal hypoglycemia (NH) after exercise is a major challenge for the management of type 1 diabetes (T1D), particularly in paediatric populations, where it is associated with significant adverse outcomes and increased carer burden. The purpose of the study was to find an interpretable early signature of the risk of post-exercise NH in individuals with T1D using CGM-derived digital biomarkers. CGM data from 49 paediatric subjects (DirecNet cohort) were used to extract several CGM metrics in two temporal configurations: (i) Exercise + Cumulative, where features were computed over the exercise window as well as over an extended window covering the period from exercise onset to recovery (16:00–17:00 and 16:00–22:00); and (ii) Exercise + Post-exercise, where features were computed separately over two non-overlapping intervals, representing the exercise phase and the subsequent recovery phase (16:00–17:00 and 17:00–22:00). A Random Forest classifier was trained using Leave-One-Out Cross Validation with variance inflation factor (VIF) based multicollinearity filtering, minimum redundancy - maximum relevance (mRMR) feature selection and SMOTE based class balancing. The Exercise + Post-exercise configuration showed the best performance: balanced accuracy (BA) = 76.9%, F1-score = 0.71, Area Under Receiver Operating Characteristic Curve (ROC-AUC) = 0.75, surpassing the Exercise + Cumulative configuration. This result was achieved with only five features: CONGA-15_EX (short-term glucose variability during exercise) was the strongest predictor, along with below_54 and above_250 (time spent in hypoglycemic and hyperglycaemic ranges), MAG (mean absolute glucose change), and GRADE_hypo (hypoglycemia risk score). The generalisability of the temporal framework was further confirmed by independent validation on the OhioT1DM free-living cohort, where the Exercise + Post-exercise configuration (BA = 76.3%, ROC-AUC = 0.804) outperformed the cumulative approach again. These results demonstrate that the temporal segmentation of CGM-derived features, extracted from the exercise and recovery windows, provides a more informative characterization of post-exercise glucose dynamics with respect the traditional cumulative representation, improving NH risk discrimination and supporting the development of lightweight CGM-only decision support tools for safer exercise management.