Individualized, self-supervised deep learning for blood glucose prediction

Abstract

AbstractThe current standard for monitoring blood glucose levels in diabetes patients are continuous glucose monitoring (CGM) devices, which are costly and carry the risk of complications, such as allergic reactions or skin irritations from the adhesive used to attach the CGM sensor to the skin. CGM devices are also highly visible and can thus act as a discomforting disease-marker for diabetes patients. To mitigate these issues, we develop and test a novel method that is able to predict blood glucose levels with only non-invasive predictor variables and a very small number of target variable measurements by using individualization and self-supervised deep learning. Using only a single blood glucose measurements per week, our method (6387.47 glucose-specific MSE) outperforms traditional deep learning performed with hourly measurements (8191.23 glucose-specific MSE). Across eight experiments where blood glucose measurements are more than one hour apart, our approach outperforms traditional deep learning without exception. Our findings suggest that self-supervised, individualized deep learning could provide an avenue towards alternatives to CGM devices that would be less costly, non-invasive, and thus more accessible.