Accurate prediction of exercise-induced hypo- and hyperglycemia remains a challenge in adults with type 1 diabetes. A study published in the Journal of Medical Internet Research developed and validated predictive models that leverage CGM data to predict these events in real-world conditions.

The analysis included 329 adults participating in the Type 1 Diabetes Exercise Initiative study. Participants completed 1,901 video-guided exercise sessions over four weeks while using CGM devices. Glycemic events were defined as blood glucose ≤54 mg/dL, ≤70 mg/dL, ≥200 mg/dL, or ≥250 mg/dL during or one hour after exercise.

Models integrating all four data types (clinical, CGM, dietary, and exercise characteristics) demonstrated excellent predictive accuracy, with cross-validated area under the receiver operating curve (AUROC) values ranging from 0.88 to 0.99. Importantly, CGM-only models performed equally well, showing strong calibration and resistance to data variability.

These results highlight that CGM-driven models can be easily implemented as low-burden decision support tools, enabling safer and more confident exercise participation for adults with type 1 diabetes.