Real-time physical activity detection module during sensor augmented insulin pump therapy

Individuals living with type 1 diabetes (T1D) face important challenges when engaging in physical activity (PA), as it necessitates careful management of blood glucose levels often through insulin adjustments and carbohydrate intake. Integrating PA detection into sensor augmented insulin pumps (SAP) is a promising strategy to enhance glycemic control by suggesting basal insulin reduction or carbohydrate ingestion in the critical 24 h after the PA detection. We have developed a real-time, model-based module for PA detection based solely on measured glucose levels, insulin infusion rates, and carbohydrate intake. The approach is based on the monitoring of the magnitude as well as various statistical properties of the prediction residuals, i.e., the discrepancies between actual sensor-measured glucose levels and model-predicted levels. We tested our algorithm on the Type 1 Diabetes and Exercise Initiative (T1DEXI) dataset, which includes structured sessions of aerobic, resistance, and interval exercises. In a dataset containing all three activity types, the detection approach based on the median of the prediction residuals successfully detected an average of 59% of PA instances, while keeping the false alarms to 3.5 per considered timeframe, when considering models tailored to each participant. When using a population model identified on in-silico data from the UVa/Padova T1D simulator, the approach successfully detected 62% of PAs, while keeping the false alarms to 3.6 per considered timeframe. These encouraging findings open the possibility of integrating PA detection into SAP systems without the need for additional physiological signals, thus enabling improved glucose management. • Activity makes Type 1 diabetes management difficult before, during, and after exercise. • We propose an algorithm to detect activity and improve glucose control. • Real-time activity detection uses glucose, insulin, and carbohydrate inputs. • The method tracks statistical patterns in Kalman Filter prediction errors. • On T1DEXI data, up to 62% of activity sessions were correctly identified.