JMIR Publications

ABSTRACT

Background:

Background:

Exercise is an important aspect of diabetes self-management. Patients with type 1 diabetes frequently struggle with exercise-induced hyperglycemia and hypoglycemia, decreasing their willingness to exercise.

Objective:

Objective:

We aim to build accurate and easy-to-deploy models to forecast exercise-induced glycemic events in real-world settings.

Methods:

Methods:

We analyzed free-living data from the Type 1 Diabetes Exercise Initiative (T1DEXI) study, where adults with type 1 diabetes wore a continuous glucose monitor (CGM) while performing video-guided exercises (30-minute exercises at least 6 times over 4 weeks), along with concurrent detailed phenotyping of their insulin program and diet. We built models to predict glycemic events (blood glucose ≤ 54 mg/dL, ≤ 70 mg/dL, ≥ 200 mg/dL, and ≥ 250 mg/dL) during and 1-hour post-exercise with variables from four data modalities: demographic and clinical; CGM; carbohydrate intake and insulin administration; and exercise type, duration and intensity.

Results:

Results:

Models incorporating information from all four data modalities showed excellent predictive performance with AUCs > 0.880 for all glycemic events. Models built with CGM data alone have statistically indistinguishable performance compared to models using all data modalities. These models also showed outstanding calibration (Brier score ≤ 0.08) and resilience to noisy input.

Conclusions:

Conclusion: We successfully constructed models to forecast exercise-induced glycemic events using only automatically captured CGM data as input, incurring minimal user burden. These models showed excellent predictive performance, calibration, and robustness, enabling model translation into a decision support tool that is easy to deploy and maintain.