Towards Wearable-based Hypoglycemia Detection and Warning in Diabetes
Martin Maritsch, Simon Föll, Vera Lehmann, Caterina Bérubé, Mathias Kraus, Stefan Feuerriegel, Tobias Kowatsch, Thomas Züger, Christoph Stettler, Elgar Fleisch, Felix Wortmann
Abstract
Rigorous blood glucose management is vital for individuals with diabetes to prevent states of too low blood glucose (hypoglycemia). While there are continuous glucose monitors available, they are expensive and not available for many patients. Related work suggests a correlation between the blood glucose level and physiological measures, such as heart rate variability. We therefore propose a machine learning model to detect hypoglycemia on basis of data from smartwatch sensors gathered in a proof-of-concept study. In further work, we want to integrate our model in wearables and warn individuals with diabetes of possible hypoglycemia. However, presenting just the detection output alone might be confusing to a patient especially if it is a false positive result. We thus use SHAP (SHapley Additive exPlanations) values for feature attribution and a method for subsequently explaining the model decision in a comprehensible way on smartwatches.