Predicting Exercise-Related Changes in Glucose in People with Type 1 Diabetes Using Linear Models and Incorporating Knowledge of Prior Exercise

Exercise-related hypoglycemia in T1D is frequent and can be life-threatening. The ability to predict glucose changes at the start of exercise could significantly improve exercise outcomes for people with T1D. We developed multivariate adaptive regression splines (MARS) to predict changes to blood glucose dynamics and to assess whether including past glucose changes during exercise improves prediction accuracy. Data was collected from 20 adults during a 4-arm, 4-day, outpatient artificial pancreas (AP) study (14 F, weight: 76.3 ±14.6 kg, age: 35.2 ± 4.6 years, height: 172.0 ± 10.6 cm). Days 1 and 4 of the study consisted of identical controlled meals and aerobic exercise events at an academic medical center. Insulin delivery data, heart rate, continuous glucose measurement (CGM) data, and carbohydrate intake were recorded. Subjects exercised at 60% of their maximal VO₂ for 45 minutes. Using the 160 aerobic exercise sessions, 2 linear predictive models were developed to estimate glucose drop during exercise: (1) a model excluding past exercise history (naïve), and (2) a model including past exercise (historical). A greedy approach was used to eliminate non-predictive features and build sparse models consisting of CGM at the start of exercise, the rate of change of CGM within 5 minutes of exercise, the metabolic expenditure 5 minutes into exercise, and participant height. Leave-one-out cross validation was done to fit the model and estimate root-mean squared error (RMSE). R² was calculated between empirical and predicted glucose drops. The naïve model was able to predict glucose drop to within 18.7 mg/dL. Including historical exercise data improved accuracy and increased R² only nominally (naïve R²=0.78, RMSE=18.7, historical R²= 0.81, RMSE=17.54). Results demonstrate prediction accuracy of linear models in estimating exercise-related glucose changes with moderate improvement by including prior exercise data.