Profiled Glucose Forecasting using Genetic Programming and Clustering

This paper proposes a method to obtain accurate forecastings of the subcutaneous glucose values from diabetic patients. Statistical techniques are applied to identify everyday situations of glucose behaviors and discover glucose profiles. This knowledge is used to create predictive models with genetic programming. The time series of glucose values, measured using continuous glucose monitoring systems, are divided into 4-hour, non-overlapping slots and clustered using a technique based on decision trees called chi-square automatic interaction detection. The glucose profiles are classified using the decision variables in order to customize the models for different profiles. Genetic programming models created with glucose values from the original dataset are compared to those of models created with classified glucose values. Significant differences and associations are observed between the glucose profiles. In general, using profiled glucose models improves the accuracy of the predictions with respect to those of models created with the original dataset.