Glucose forecasting using genetic programming and latent glucose variability features

Abstract

This paper investigates a set of genetic programming methods to obtain accurate predictions of subcutaneous glucose values from diabetic patients. We explore the usefulness of different features that identify the latent glucose variability. New features, including average glucose, glucose variability and glycemic risk, are generated as input variables of the genetic programming algorithm in order to improve the accuracy of the models in the prediction phase. The performance of traditional genetic programming, and models created with classified glucose values, are compared to those using latent glucose variability features. We experimented with a set of 18 different features and we also performed a study of the importance of the variables in the models. The Bayesian statistical analysis shows that the use of glucose variability as latent variables improved the predictions of the models, not only in terms of RMSE, but also in the reduction of dangerous predictions, i.e., those predictions that could lead to wrong decisions in the clinical practice.