A simple method for recommending specialized specifications for diabetes monitoring

Abstract

Under glycemic variability, a characterization of the desired blood glucose (BG) behavior is needed to assess if a given artificial pancreas (AP) respects its specification. The specification is an essential element to detect any deviation from an adequate insulin policy. Specializing the monitoring specification is therefore of utmost importance as existing guidelines for diabetes management are general and do not take into account how the personal factors and lifestyle affect the glycemic behavior. Surely, recommending personalized monitoring specifications may provide flexible and appropriate treatment goals to be attained by diabetic patients in order to account for their actual treatment needs. In this work, we use machine learning models to characterize glycemic behavior in synthetic healthy individuals. To account for the day-by-day fluctuation in BG levels, we use a stochastic process superimposed on a deterministic model of the glucose-insulin dynamics. The obtained characterization of the glycemic behavior in healthy individuals is then used as the target class to predict, and thus recommend, personalized monitoring specifications to diabetic patients. Results show that the approach stands as a feasible strategy to recommending appropriate and realistic monitoring goals for diabetic patients based on healthy individuals who share a similar glycemic behavior. Eventually, the incorporation of a recommender approach on an intelligent monitoring system for the AP will allow on-line adaptation of the treatment requirements for each patient.