Automatic Detection of Exercise in People with Type 1 Diabetes Using an Unscented Kalman Filter [Pòster]

Abstract

Physical activity in type 1 diabetes mellitus has been found to have varying degrees of effect on glycaemic control depending on the type, intensity and duration of the exercise. Such effect is associated with an imbalance between hepatic glucose production and glucose disposal into the muscle, increased insulin sensitivity and impaired counter-regulatory hormonal response. In the context of an artificial pancreas, automatic detection of exercise has the potential to significantly improve glycaemic control. This work aims to develop a new methodology for automatically detecting exercise that only requires data from a continuous glucose monitor (CGM) and the insulin delivered to the subject. Method: The glucose-insulin Minimal Model was extended, by adding an insulin absorption model and an auxiliary parameter used to describe disturbances. Increases in this parameter may indicate meal ingestions whereas decreases may indicate exercise. The disturbance parameter was estimated using an Unscented Kaman Filter. Two thresholds were introduced to detect exercise: a first threshold to indicate the possibility of an abnormal event; and a second threshold, based on an area-under-the-curve, to indicate exercise. The method was tested on data from 7 closed-loop trials including a period of moderate intensity structured exercise. Results: Overall, the results obtained were satisfactory with an average detection time of 22 minutes, accuracy of 96%, sensitivity of 100% and a specificity of 96%. Conclusion: The presented technique has the potential to be a viable approach to detect physical exercise in the context of an artificial pancreas. Improvements and further testing are necessary to confirm such hypothesis (Diabetes Technology & Therapeutics, 2016, vol. 18, núm. S1, https://doi.org/10.1089/dia.2016.2525)