(184s) A Monte Carlo Simulation Study to Evaluate the Limits of Prediction Accuracy for Blood Glucose Concentration

Predictions of blood glucose concentration is an important aspect of glucose management for people with type 1 diabetes. Especially in applications of artificial pancreas (AP), systems that can automatically determine insulin infusion rate with the frequency up to every minute, model-based control algorithms utilize predictions of glucose extensively. Model predictive control (MPC), as one of the most widely used control algorithms in this application, requires long prediction horizons into the future (e.g. over 5 hours into the future) due to long inputs dynamics (i.e. carbs and insulin).

This study used data collected from 11 subjects with type 1 diabetes under free-living conditions (i.e. no restrictions on the life styles of participating subjects) with noises added to mimic sensor noise/measurement errors during the study. The autoregressive (AR) models, widely used in AP applications, were used to model predictions of glucose up to one (1) hour into the future. First part of the study was to examine the deterioration of prediction accuracy as predication horizon increases. Then the effects of noises were evaluated. Lastly, an adaptive modeling methodology was evaluated against fixed model parameterization.

In terms of the average absolute deviation (AAD), indicating the model deviations, the model performance deteriorates almost linearly as prediction horizon increases. The AAD reached over 35 mg/dL, and the correlation coefficient between measured (treated as true glucose concentrations) and predicted glucose (rfit) fell to 0.74, with prediction of one hour into the future. The AAD also increases as the noise added increases. However, as standard deviation of added white noise increases from 5 to 15 mg/dL, the AAD for one hour prediction horizon only increased by 1.1%. For the adaptive modeling methodology, the AAD and rfit improved from mean AAD of 34.41 mg/dL to 33.32 mg/dL against fixed model parameterization.

The study shows prediction accuracy deteriorates mainly due to long prediction horizons other than sensor noises. Additionally, the adaptive modeling methodology improves model performance. However, the limits of prediction accuracy are mainly bound by the length of prediction horizons. Control algorithms that avoid the use of long prediction horizons may be preferable options for this AP application.