A Multiple Local Models Approach to Accuracy Improvement in Continuous Glucose Monitoring

Background: Continuous glucose monitoring (CGM) devices estimate plasma glucose (PG) from measurements in compartments alternative to blood. The accuracy of currently available CGM is yet unsatisfactory and may depend on the implemented calibration algorithms, which do not compensate adequately for the differences of glucose dynamics between the compartments. Here we propose and validate an innovative calibration algorithm for the improvement of CGM performance. Methods: CGM data from GlucoDay (R) (A. Menarini, Florence, Italy) and paired reference PG have been obtained from eight subjects without diabetes during eu-, hypo-, and hyperglycemic hyperinsulinemic clamps. A calibration algorithm based on a dynamic global model (GM) of the relationship between PG and CGM in the interstitial space has been obtained. The GM is composed by independent local models (LMs) weighted and added. LMs are defined by a combination of inputs from the CGM and by a validity function, so that each LM represents to a variable extent a different metabolic condition and/or sensor-subject interaction. The inputs best suited for glucose estimation were the sensor current I and glucose estimations (G) over cap, at different time instants [I-k, Ik-1, (G) over cap (k-1)] (IIG). In addition to IIG, other inputs have been used to obtain the GM, achieving different configurations of the calibration algorithm. Results: Even in its simplest configuration considering only IIG, the new calibration algorithm improved the accuracy of the estimations compared with the manufacturer's estimate: mean absolute relative difference (MARD) = 10.8 +/- 1.5% versus 14.7 +/- 5.4%, respectively (P = 0.012, by analysis of variance). When additional exogenous signals were considered, the MARD improved further (7.8 +/- 2.6%, P<0.05). Conclusions: The LM technique allows for the identification of intercompartmental glucose dynamics. Inclusion of these dynamics into the calibration algorithm improves the accuracy of PG estimations.