Online Classification of Unstructured Free-Living Exercise Sessions in People with Type 1 Diabetes

Area of Science:

• Endocrinology and Metabolism
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Exercise management in type 1 diabetes (T1D) is challenging due to diverse glycemic impacts.
• Automated insulin delivery (AID) systems require accurate exercise classification for optimal glucose control.

Purpose of the Study:

• To develop and validate a classification model for exercise events (structured and unstructured) into aerobic, interval, or resistance types.
• To enable integration into AID systems for managing exercise-induced glycemic disturbances.

Main Methods:

• A long short-term memory network model was trained using real-world data from structured exercise sessions.
• Data included triaxial accelerometer, heart rate, and activity duration.
• The algorithm classified 15 common free-living activities and related them to glucose changes.

Main Results:

• The model achieved accuracies of 72% for aerobic, 65% for interval, and 77% for resistance exercise in structured sessions.
• Tested on 3328 unstructured sessions, the algorithm identified physical activity associated with expected glucose changes.
• Mean glucose changes were -20.8 (aerobic), -16.2 (interval), and -11.6 (resistance) mg/dL.

Conclusions:

• The developed algorithm reliably identifies physical activity and its associated glucose changes.
• This classification capability can be integrated into AID systems to personalize glycemic management during exercise.