Exercise can be challenging for people with type 1 diabetes because exercise-induced hypoglycemia during exercise is common. The problem is compounded because different types of exercise can cause different responses. Anaerobic exercise can cause less of a drop in glucose than aerobic. And interval exercise can actually cause glucose to increase. People respond differently to exercise at different times of day with differing insulin on board.
We present data across multiple studies that include CGM, insulin, food, and physical activity metrics in people with T1D during different types of exercise including aerobic, anaerobic, and interval exercises and across various insulin infusion therapies including single and dual-hormone closed loop, sensor-augmented pump, and multiple-daily-injections. In addition to free-living data, we have data on the same subjects, using identical therapies, performing identical exercise at the same time of day to explore the limits of predictability under ideal, repeated circumstances. We present the design of various glucose prediction machine learning algorithms including random forests, multivariate adaptive regression splines (MARS), neural networks, and support vector regression algorithms.
Results indicate the importance of including heterogeneous training data across multiple exercise modalities, insulin therapies, and times of day when building predictive models. Prediction accuracy was not found to improve substantially when including data from prior exercise sessions. CGM features were the most relevant features in improving prediction accuracy.
Predicting glucose changes during exercise requires algorithms that have been trained on large heterogeneous datasets with observations from different exercise modalities, insulin therapies, and times of day.