Sparse Representation Models of Continuous Glucose Monitoring Time-Series

Abstract

Continuous glucose monitoring (CGM) is essential towards the effective management of type 1 diabetes. Reliable CGM time-series models can afford new insights into treatment by allowing us to identify clinically-meaningful signal components and rule-out noise. We propose the use of sparse representation techniques with appropriately designed dictionaries to express CGM signals as a linear combination of a small set of knowledge-driven atoms. Our results indicate that the proposed framework consists a viable solution for modeling CGM time-series reaching relative reconstruction errors of 0.08 and suggest that this approach can be used to interpret the underlying CGM time-series in relation to clinical assessments.